Pith. sign in

REVIEW 3 major objections 5 minor 83 references

Behavior cloning finetuning erases what a vision-language model knows; this paper shows that distilling from a frozen copy and aligning language with motion-direction labels preserves it and nearly doubles real-robot success.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 05:11 UTC pith:AAJ5LLTZ

load-bearing objection Solid recipe paper: Anchor-Align's two auxiliary losses give consistent gains across sim and real, but the fixed loss weights and a partly circular diagnostic need scrutiny before betting on the exact magnitudes. the 3 major comments →

arxiv 2607.13429 v1 pith:AAJ5LLTZ submitted 2026-07-15 cs.RO cs.CV

Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment

Dwip Dalal , Shivansh Patel , Chahit Jain , Jeonghwan Kim , Utkarsh Mishra , Alex Baratian , Hyeonjeong Ha , Heng Ji
show 2 more authors
Svetlana Lazebnik Unnat Jain
This is my paper
classification cs.RO cs.CV
keywords vision-language-action modelsbehavior cloningcatastrophic forgettingrepresentation distillationlanguage-action alignmentout-of-distribution generalizationrobot manipulationfinetuning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard finetuning of a pretrained vision-language model into a robot policy by behavior cloning progressively overwrites the visual and semantic representations that made the model useful, so the policy memorizes the training scene instead of following new instructions. This paper proposes Anchor-Align, two additions to the cloning loss: layer-wise distillation of hidden states from a frozen copy of the pretrained model, and a cross-entropy loss that makes the language head predict the robot's motion direction on the same observation the action head acts on. The claim is that these two terms together keep the pretrained semantics intact and ground language in action, with no extra data or human annotation. The reported consequence is roughly doubled real-robot success and large gains on out-of-distribution and long-horizon benchmarks across two architecturally different VLA setups; the authors also show co-training on image-text data neither prevents forgetting nor aligns the language and action heads.

Core claim

Anchor-Align's central claim is that VLA finetuning fails for two separable reasons—unconstrained behavior cloning overwrites pretrained vision-language representations, and co-training supervises language and action on different observations so the two heads can contradict each other—and that both are fixable by reusing supervision already present in the demonstrations. Vision-Language Anchoring adds a per-layer squared-error loss between the trainable backbone's hidden states and those of a frozen copy of the same pretrained VLM, evaluated on every decoder layer and every vision and text token. Language-Action Alignment converts each action chunk into one of six direction words by averagin

What carries the argument

The load-bearing object is the combined loss L = L_action + λ_anchor L_anchor + λ_align L_align. L_anchor is the layer-wise squared Frobenius difference between the hidden states of the trainable backbone and a frozen copy of the pretrained VLM, averaged over all decoder layers; it is what stops the backbone from drifting. L_align is a six-way cross-entropy over motion-direction words (up/down/left/right/forward/backward) derived programmatically from each action chunk by averaging, filtering near-stationary chunks, and selecting the dominant axis and sign; it is applied to the pre-action hidden state of the last instruction token and projected through the frozen language-model head. The act

Load-bearing premise

The load-bearing premise is that a weighted layer-wise pull toward a frozen VLM copy (λ_anchor = 0.1) preserves pretrained semantics without blocking robot-specific adaptation, and that six coarse motion-direction words capture enough of the action to align language with behavior rather than regularizing away control signal.

What would settle it

Sweep λ_anchor from 0 to 1 on a fixed benchmark while measuring both task success and preserved visual-reasoning accuracy; if success peaks only at exactly 0.1 and collapses nearby, or if removing the alignment loss leaves the perceptual-robustness benchmark unchanged, the central causal claim fails. A second check: evaluate the real-robot semantic test with language labels that are action-consistent but not motion-direction words; if the gains vanish, the specific alignment mechanism, not generic language supervision, is what matters.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the mechanism is right, standard BC's out-of-distribution failures are not irreducible: position-swap success rises from 2.3% to 22.6% with no new data.
  • Co-training on web image-text data is not the remedy for forgetting; the co-trained baseline scores near zero on position swap while anchoring retains roughly 70% of pretrained visual-reasoning accuracy.
  • Preserving pretrained representations and learning control are compatible, not opposed; the method reaches the highest action decodability while sustaining high language preservation.
  • Language-action alignment is causally tied to success: shuffled or meaningless direction labels collapse to baseline, and per-rollout alignment correlates positively with task success.
  • The recipe transfers across architectures and action heads: both a regression head and a flow-matching diffusion head show large real-robot gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same two mechanisms suggest a general recipe for any VLM-to-policy adaptation—anchor the backbone to its pretrained copy and co-supervise the language channel with action-derived labels; this could be tested on other robot embodiments and other foundation models.
  • Editorial inference: because the paper never sweeps the anchoring weight, the claim that 0.1 is the right balance between preservation and adaptation is untested; a sweep would reveal whether the gains are a narrow tuning artifact or a broad plateau.
  • Editorial inference: the six-word vocabulary is coarse by construction; a finer discretization, or continuous language supervision, could either strengthen grounding or over-regularize control, and the paper's own diagnostic framework makes that question directly measurable.
  • Editorial inference: the alignment measure itself—frame-level agreement between language and discretized action—could serve as a cheap training-time proxy for out-of-distribution generalization in future VLA work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Anchor-Align, a VLA finetuning recipe that adds two objectives to the standard behavior-cloning loss: (1) Vision-Language Anchoring, a layer-wise distillation of backbone hidden states from a frozen copy of the pretrained VLM, and (2) Language-Action Alignment, a cross-entropy loss that supervises a language-head prediction of a six-way motion-direction label derived programmatically from the continuous action chunk on the same observation. The method is evaluated on LIBERO-PRO, LIBERO-Plus, CALVIN ABC→D, and a physical xArm7 platform, using VLA-Adapter and StarVLA backbones. The central claim is that preserving and aligning pretrained VLM representations during action finetuning improves OOD spatial generalization, perceptual robustness, long-horizon control, and real-robot success without extra data or annotation.

Significance. If the claims hold, this is a practically valuable result: a simple, low-overhead finetuning recipe that improves VLA generalization on multiple external benchmarks and two real-robot action-head architectures. The paper has several genuine strengths: the Shuffle and Scatter controls in Table 4 go beyond typical auxiliary-loss ablations and directly address the 'any auxiliary loss helps' alternative; Appendix C.3 provides five-seed means for the main method and its closest baseline on key axes; the real-robot experiments cover two architecturally distinct backbones; and the training-cost comparison in Table 6 shows the method is cheap relative to co-training. The central idea—that a frozen-teacher distillation loss plus a coarse language-action consistency loss can mitigate catastrophic forgetting without blocking action learning—is plausible and well motivated. However, the evidence for the specific operating point is incomplete: the two loss weights and the stationarity threshold are fixed without a sensitivity analysis, and the language-action alignment diagnostic in Section 4.5 reuses the exact discretization used to generate the training labels, which weakens the mechanistic i

major comments (3)
  1. [§3.2–3.3, Eqs. (1)–(5), App. B.2] The three hyperparameters λ_anchor=0.1, λ_align=0.02, and τ=0.15 are fixed and no sensitivity study is reported. This matters because L_anchor in Eq. (2) is a sum of per-layer squared Frobenius norms over N×d entries, where N is the number of vision and text tokens; the gradient scale of L_anchor relative to L_action is therefore not set by λ_anchor alone and can vary with sequence length and vision-patch count. The frozen-backbone comparison (43.1% on LIBERO-PRO) brackets the extreme end of the preservation–plasticity trade-off but does not show how close λ_anchor=0.1 is to a stability cliff. I request a sweep over λ_anchor (e.g., 0.01–1), λ_align, and τ on at least LIBERO-PRO position-swap/object-swap and LIBERO-Plus, together with reported gradient norms of the two auxiliary losses relative to the action loss.
  2. [§4.5, Table 5] The alignment diagnostic uses the same dominant-axis discretization and the same six-word vocabulary that generate the L_align training labels in Sec. 3.3. Comparing the language head's prediction with the policy action discretized into that same label space risks measuring the model's ability to reproduce a training label rather than an independent behavioral consistency. The strong Pearson correlation (r=0.51) between per-rollout alignment and success is interesting, but it should be validated with held-out annotations or a different label axis (e.g., orientation or grasp labels not used during training). At minimum, the paper should state and discuss this circularity explicitly and show alignment on rollouts where the discretization was not used to supervise L_align.
  3. [Tables 1–2 and Fig. 7] Several headline results are reported from single seeds or from a small number of rollouts. Appendix C.3 gives five-seed means only for Anchor-Align and VLA-Adapter on selected LIBERO axes, not for all baselines in Table 1, and Table 2 (CALVIN) is single-seed. The real-world results in Fig. 7 are based on 20 rollouts per condition, with no confidence intervals or per-condition significance tests. Given the paper's claim of 'consistent improvements' across all benchmarks, I ask for multi-seed results or confidence intervals at least for the central Anchor-Align vs. VLA-Adapter comparison, and for the CALVIN and real-robot headline numbers.
minor comments (5)
  1. [App. C.3] The text refers to 'the main paper's Tab. 7', but Table 7 is in the appendix; renumber or re-reference. Also, the text says 'mean and variance' while Table 8 reports mean ± standard deviation.
  2. [Eq. (2)] The notation H_S^u[m] and H_A^u[m] should clarify whether the norm is averaged over unmasked token positions or over all N positions, and whether padding or mask-dependent token counts affect the scale of L_anchor.
  3. [Fig. 16 and App. E.2] There are typographical artifacts: 'lib' appears instead of 'lid' and several Unicode soft-hyphen sequences appear as 'uni00A0' in the appendix figures and text.
  4. [Sec. 3.3] The tie-breaking rule in the direction discretization (x > y > z) is stated only in the appendix, but it is part of the method definition and should appear in the main text or be referenced clearly.
  5. [Sec. 4.4] Fig. 7 labels the mean with '(± SE)', but no error bars are shown; either add error bars or remove the parenthetical claim.

Circularity Check

2 steps flagged

Benchmark gains are independent, but two internal diagnostics are defined by the same labels/representations the training losses optimize.

specific steps
  1. self definitional [Sec. 3.3 (Eq. 5) and Sec. 4.5]
    "we programmatically convert the continuous action target acont into a motion-direction label alang ... and use this label to supervise the VLM backbone. ... We measure this misalignment by extending the programmatic framework of Sec. 3.3, which converts robot demonstrations into language labels, into a diagnostic tool: at each timestep, we compare the model's language prediction with its discretized policy action."

    The L_align objective in Eq. (5) is CE(âlang, alang), where alang is produced by the dominant-axis discretization in Sec. 3.3. The Sec. 4.5 'alignment' metric is the agreement between the language head and the same discretized action label. For Anchor-Align, the language head is directly trained to output alang, and the action head is trained by BC on the same ground-truth actions from which alang is derived. Thus the reported 78.4% alignment is substantially a check that the model reproduces its own training labels, not an independent measurement of cross-head consistency. The diagnostic is informative for co-trained baselines that never saw these labels, but the Anchor-Align number is forced by construction.

  2. self definitional [App. D.3 and Eq. (2)]
    "Language preservation, measured with Centered Kernel Alignment (CKA) between the finetuned backbone's text-token hidden states and the pretrained VLM's. ... The per-layer anchoring loss is L_anchor^(u) = ||H_S_u[m] - H_A_u[m]||_F^2."

    L_anchor directly minimizes the squared Frobenius distance between the trainable backbone's hidden states and the frozen pretrained VLM's hidden states at every decoder layer. CKA measures geometric similarity between the same two sets of hidden states. Anchor-Align's high CKA (0.91 overall, 0.95 average) is therefore a direct consequence of optimizing the anchoring objective, not an emergent property. The paper's non-circular evidence for preservation is the behavioral GQA curve in Fig. 9, which is measured on an external benchmark; the CKA diagnostic itself is self-definitional.

full rationale

The central claims—LIBERO-PRO/LIBERO-Plus success, CALVIN long-horizon completion, and real-robot gains—are measured on external benchmarks and are not defined by the training losses, so the main derivation is independent. The circularity is confined to two internal diagnostics: the language-action alignment metric reuses the exact discretization/label pipeline that generates L_align, and the CKA language-preservation metric measures the same hidden-state distance that L_anchor explicitly minimizes. These make the reported alignment and CKA numbers partially checks that the model fits its own training targets rather than independent discoveries. The GQA curve and the Shuffle/Scatter controls provide some non-circular support. Self-citations are present but are background/related-work citations and are not load-bearing premises. Weighing the self-referential diagnostics against the independent benchmark results, a score of 4 is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

Anchor-Align adds two loss terms with hand-set weights, a stationarity threshold, and a hand-chosen six-word label space; these are free parameters for the reported results. The central assumptions are that a pretrained VLM's features are worth preserving via layer-wise distillation, that coarse dominant-axis direction words provide useful language-action supervision, and that the public OOD benchmarks used here are valid proxies. No new physical entities are introduced; the anchor teacher is the same pretrained VLM, so the method adds no external knowledge beyond the original checkpoint.

free parameters (5)
  • lambda_anchor = 0.1
    Weight of layer-wise anchoring loss in Eq. (1); set without sensitivity analysis in App. B.2, could be tuned to LIBERO validation.
  • lambda_align = 0.02
    Weight of language-action alignment loss in Eq. (1); set without sensitivity analysis.
  • tau (stationarity threshold) = 0.15
    Filtering threshold for near-stationary samples when building direction labels; affects which frames get L_align supervision.
  • direction label set = six words: left, right, up, down, forward, backward
    Hand-chosen label space for discretizing continuous actions; arbitrary and may not capture diagonal or combined motions.
  • LoRA rank/alpha = r=64, alpha=128
    PEFT configuration chosen from VLA-Adapter defaults, not independently justified.
axioms (6)
  • domain assumption Pretrained Qwen2.5-0.5B contains transferable visuospatial and semantic priors.
    Central premise of the anchoring objective; invoked throughout Sec. 3.2 and Sec. 4.2.
  • domain assumption Layer-wise feature matching to a frozen copy preserves representations without blocking adaptation.
    The paper's main mechanism; Eq. (2) is assumed to be a harmless regularizer, but the frozen-backbone baseline in App. D.3 shows too much preservation hurts.
  • domain assumption Dominant-axis discretization of action chunks into six direction words is valid supervision.
    Sec. 3.3 derives language labels from averaged action targets; this assumes the dominant translational axis is the right semantic abstraction.
  • domain assumption Frozen LM head remains a good classifier for direction words after LoRA finetuning.
    Eq. (4) routes the pre-action hidden state through the frozen W_lm; assumes the pretrained output embedding is compatible with the new direction-word vocabulary.
  • domain assumption Benchmark OOD perturbations in LIBERO-PRO/Plus/CALVIN are representative of real-world generalization.
    The simulation claims rely on these benchmarks as proxies; the paper also validates on a physical xArm7, but with limited scale.
  • ad hoc to paper Language-action alignment measured by comparing language and discretized action heads reflects meaningful behavioral consistency.
    Sec. 4.5 introduces this diagnostic; it shares its discretization procedure with the L_align training loss, so its validity as an independent metric is partly self-referential.

pith-pipeline@v1.3.0-alltime-deepseek · 32165 in / 15554 out tokens · 145203 ms · 2026-08-02T05:11:39.806606+00:00 · methodology

0 comments
read the original abstract

Finetuning a pretrained vision-language model (VLM) on robot demonstrations via behavior cloning (BC) has become the standard recipe for vision-language-action (VLA) policies. However, BC finetuning progressively overwrites the pretrained representations that support visual and semantic generalization. Co-training on web image-text data, a common remedy, does not prevent this; it applies language and action losses to separate observations, leaving VLAs with language-action misalignment that standard manipulation benchmarks do not expose. We propose Anchor-Align, which augments BC with two objectives: Vision-Language Anchoring distills layer-wise representations from a frozen VLM copy to prevent this drift, while Language-Action Alignment converts each action target into a discrete motion-direction label and jointly trains language and action prediction on the same robot observation. On a physical xArm7 robot, across two widely used VLA architectures, Anchor-Align improves real-robot success on both (28% to 54% and 37% to 60%). At scale in simulation, we demonstrate consistent improvements on OOD perturbations, perceptual robustness, and long-horizon control across LIBERO-PRO, LIBERO-Plus, and CALVIN, respectively, suggesting that preserving pretrained representations and effective action learning are not fundamentally at odds. Project page: anchoralignvla.github.io

Figures

Figures reproduced from arXiv: 2607.13429 by Alex Baratian, Chahit Jain, Dwip Dalal, Heng Ji, Hyeonjeong Ha, Jeonghwan Kim, Shivansh Patel, Svetlana Lazebnik, Unnat Jain, Utkarsh Mishra.

Figure 1
Figure 1. Figure 1: Out-of-distribution test for VLAs. Standard BC, finetuned to pick up the green mug, reaches for the green mug even when instructed to pick up the pink mug, whereas Anchor-Align retains the VLM’s representations and successfully completes the task. GT Text Anchor Loss GT Action BC Loss Anchor VLM VLM Backbone Ln L1 Action Expert Proprioceptive State & Initial Action Ln L1 . . . . . . . . . . . . . . . D1 Dn… view at source ↗
Figure 2
Figure 2. Figure 2: Anchor-Align preserves VLM priors and aligns them with action during VLA finetuning. Vision-Language Anchoring distills representations from a frozen pretrained VLM into the trainable VLA at every transformer layer (Anchor loss), preventing catastrophic forgetting. Language-Action Alignment pro￾grammatically converts each ground-truth demonstration trajectory into a discrete language label and trains the m… view at source ↗
Figure 3
Figure 3. Figure 3: Anchor-Align VLA aligns the VLM’s language predictions with action generation during VLA finetuning. Standard BC supervises only action prediction, allowing the model’s language output (red) to contradict the required motion on the same observation (blue). Language-Action Alignment derives a discrete motion-direction label from each ground-truth action target and trains the backbone to predict it on the sa… view at source ↗
Figure 4
Figure 4. Figure 4: Anchor-Align VLA generalizes to semantic perturbations. Both rows are from LIBERO-PRO: the top shows the position-swap axis and the bottom the object-swap axis. Standard BC fails by executing trajectories tied to the training scene, such as reaching toward the original object location or missing the resized target. Anchor-Align VLA grounds to the perturbed observation and succeeds. Anchor-Align improves ro… view at source ↗
Figure 5
Figure 5. Figure 5: Real-world rollouts: Anchor-Align generalizes across spatial setups, whereas standard BC fails. Each row is a held-out perturbation regime (compositional object layout, spatial rearrangement, and se￾mantic perturbation); the layout seen during training (left) always differs from the one at test (right). Setup 1 and Setup 2 are two distinct spatial configurations of the same task, with the target, objects, … view at source ↗
Figure 6
Figure 6. Figure 6: Cluttered-scene. Starting from a heavily cluttered tabletop (left), Anchor-Align VLA picks the language-specified target across many distractors, grounding each instruction in the current scene. Compositional Object￾Layout Spatial Rearrange￾ment Cluttered Scene Mean (± SE) 20 40 60 80 100 Success Rate (%) 20% 22.5% 42.5% 35% 28.3% 42.5% 85% 54.2% Backbone: Prismatic-Qwen2.5-0.5B & Bridge Attention Head Sta… view at source ↗
Figure 7
Figure 7. Figure 7: Real-world generalization across different perturbations and backbones. (a, b) Success rates on three generalization tests (spatial rearrangement, cluttered scene, and compositional object-layout) for two VLA backbones (20 rollouts per condition). Anchor-Align improves every condition on both backbones, so the gains are not tied to a specific backbone or action head. (c) A pink-mug semantic-perturbation te… view at source ↗
Figure 8
Figure 8. Figure 8: Failure mode analysis comparing standard BC and Anchor-Align. Limitations and Error Analysis. We man￾ually inspect all failed real-world rollouts and bin clearly attributable failures into six modes ( [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual reasoning accuracy dur￾ing finetuning. Standard BC loses 94% of its GQA accuracy within 10K steps (orange), while Anchor-Align retains 70% (green); the frozen VLM is the upper bound (dashed). For effective generalization, a VLA must be able to per￾form visual reasoning, i.e., understand objects, their semantics, and their attributes in a scene, an ability that is a hallmark of pretrained VLM represe… view at source ↗
Figure 10
Figure 10. Figure 10: LIBERO Long Suite. Performance across nine evaluation axes (excluding Standard and Position Swap). Anchor-Align VLA (orange) substantially expands coverage over VLA-Adapter (gray), with the largest gains on Lighting Condition (+20.8%), Object Layout (+18.6%), and Camera Viewpoint (+17.7%). C.2 Per-Suite Robustness Breakdowns In the main paper (Tab. 1), we report robustness and generalization results on th… view at source ↗
Figure 11
Figure 11. Figure 11: LIBERO Object Suite. Performance across nine evaluation axes (excluding Standard and Position Swap). Anchor-Align VLA (orange) consistently covers a larger area than VLA-Adapter (gray), with the most pronounced gains on Robot Init State (+18.6%). Language Rephrase Object Swap Language Instruct. Background Texture Robot Init State Camera Viewpoint Object Layout Lighting Condition Sensor Noise 50% 60% 70% 8… view at source ↗
Figure 12
Figure 12. Figure 12: LIBERO Goal Suite. Performance across nine evaluation axes (excluding Standard and Position Swap). Anchor-Align VLA (orange) shows the largest gains over VLA-Adapter (gray) on Language Instruction (+11.9%) and Background Texture (+4.3%). On LIBERO-PRO, Anchor-Align VLA improves the StarVLA baseline from 73.6% to 89.0% on the language rephrase axis and from 89.8% to 91.2% on the object swap axis. The impro… view at source ↗
Figure 13
Figure 13. Figure 13: Action decodability vs. language preservation. Per-model language preservation (text-token CKA) vs. action decodability (linear￾probe R2 ): Anchor-Align VLA attains the high￾est action decodability while sustaining high lan￾guage preservation, gaining action information without overwriting the pretrained geometry. 1 4 8 12 16 20 24 VLM Transformer Layer 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Text-Token CKA (… view at source ↗
Figure 15
Figure 15. Figure 15: Illustrating all four diagnostic axes. Each robot manipulation episode is segmented into four diagnostic axes: Motion Direction (t=1 to 15), Orientation (t=34 to 41), Grasp (t=58 to 64), and Task Com￾pletion (t=83 to 91). For each axis, ground-truth labels are extracted and paired with template questions for evaluation. Ns = 10 frames are sampled uniformly at random from each stage via seeded sampling, yi… view at source ↗
Figure 16
Figure 16. Figure 16: Real-world rollout completion time. Distribution of successful rollout durations on the xArm7 pick-and-place task. Anchor-Align VLA completes rollouts 1.7× faster than standard BC, with lower variance. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Target-object orientation perturbation (Compositional Object-Layout). The same broccoli, cropped from the front-view camera, shown in six distinct orientations used across evaluation rollouts of the ”Pick up the green broccoli and place it on the plate task”. We perturb the orientation of the object itself, not just its position, so its silhouette and visible structure change substantially while its ident… view at source ↗
Figure 18
Figure 18. Figure 18: Successful real-world rollouts from the Anchor-Align VLA model. The task is to pick up the broccoli and place it on the green plate. The broccoli’s spatial position and the surrounding distractor objects are simultaneously swapped across rollouts, so each rollout is evaluated on a unique scene configu￾ration. All six rollouts succeed, showing that Anchor-Align VLA handles joint object-swap and position-sw… view at source ↗
Figure 19
Figure 19. Figure 19: Successful real-world rollouts from the Anchor-Align VLA model. The task is to pick up the green mug and place it on the plate. The mug’s spatial position and the surrounding distractor objects are simultaneously swapped across rollouts, so each rollout is evaluated on a unique scene configuration. All six rollouts succeed, showing that Anchor-Align VLA handles joint object-swap and position-swap perturba… view at source ↗
Figure 20
Figure 20. Figure 20: Successful real-world rollouts from the Anchor-Align VLA model. The task is to pick up the object named by the language instruction (pineapple, red bell pepper, or yellow bell pepper) and place it on the plate in a very cluttered scene. The target is specified by text while many distractor objects (including the other candidate items) crowd the workspace, so the policy must ground the instruction in the c… view at source ↗
Figure 21
Figure 21. Figure 21: LIBERO-Goal Plus: Background and Robot perturbations. Top: Background perturbation on “Put the bowl on the plate”—the table texture is changed to stone tile. Standard BC (red) collapses onto the table without releasing the bowl on the plate; Anchor-Align VLA (green) places the bowl on the plate. Bottom: Robot perturbation on “Open the middle drawer of the cabinet”—the robot’s initial joint configuration i… view at source ↗
Figure 22
Figure 22. Figure 22: LIBERO-Goal Plus: Objects and Background perturbations. Top: Objects perturbation on “Open the middle drawer of the cabinet”—the cabinet, plate, and surrounding distractors are rearranged. Stan￾dard BC (red) hovers above the wrong region without contacting the handle; Anchor-Align VLA (green) locates the middle drawer and opens it. Bottom: Background perturbation on “Put the bowl on top of the cabinet”—th… view at source ↗
Figure 23
Figure 23. Figure 23: LIBERO-Long Plus: Camera and Language perturbations. Top: Camera perturbation on the long-horizon task “Put both the cream cheese box and the butter in the basket”: the camera viewpoint is shifted. Standard BC (red) picks up the incorrect object and places it in the basket, showcasing its reliance on memorized scene-to-action mappings that break under the viewpoint shift; Anchor-Align VLA (green) sequenti… view at source ↗
Figure 24
Figure 24. Figure 24: LIBERO-Long Plus: Lighting and Camera perturbations. Top: Lighting perturbation on “Put both the alphabet soup and the tomato sauce in the basket”: the scene is darkened. Standard BC (red) grasps an incorrect object, places it in the basket, and then idles without ever reaching for the second target; Anchor￾Align VLA (green) deposits both items in the basket. Bottom: Camera perturbation on the long-horizo… view at source ↗
Figure 25
Figure 25. Figure 25: LIBERO-PRO Object Swap. Both rows use the same object-swap regime: the basket is replaced with a pink wicker variant and the surrounding canned items are exchanged for novel objects unseen during training. Top: “Pick up the alphabet soup and place it in the basket”: Standard BC (red) hovers over an incorrect target and then moves to the basket without grasping any object, whereas Anchor-Align VLA (green) … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

83 extracted references · 1 canonical work pages

  1. [1]

    S. Bahl, R. Mendonca, L. Chen, U. Jain, and D. Pathak. Affordances from Human Videos as a Versatile Representation for Robotics. InConference on Computer Vision and Pattern Recognition, 2023

    2023

  2. [2]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, et al. Qwen2.5-VL Technical Report.arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Bjorck, F

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, et al. GR00T N1: An Open Foundation Model for Generalist Humanoid Robots.arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, et al.π 0: A Vision-Language- Action Flow Model for General Robot Control. InRobotics: Science and Systems, 2025

    2025

  5. [5]

    Y . Bo, A. Soni, S. Srivastava, and M. Khosla. Evaluating representational similarity measures from the lens of functional correspondence.arXiv preprint arXiv:2411.14633, 2024

    arXiv 2024

  6. [6]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Haus- man, A. Herzog, J. Hsu, et al. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Q. Bu, H. Li, L. Chen, et al. Towards Synergistic, Generalized, and Efficient Dual-System for Robotic Manipulation.arXiv preprint arXiv:2410.08001, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Q. Bu, Y . Yang, J. Cai, S. Gao, G. Ren, M. Yao, P. Luo, and H. Li. Univla: Learning to act anywhere with task-centric latent actions.arXiv preprint arXiv:2505.06111, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Cadene, S

    R. Cadene, S. Alibert, A. Soare, Q. Gallouedec, A. Zouitine, S. Palma, P. Kooijmans, M. Ar- actingi, M. Shukor, D. Aubakirova, M. Russi, F. Capuano, C. Pascal, J. Choghari, J. Moss, and T. Wolf. LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch. https://github.com/huggingface/lerobot, 2024

    2024

  10. [10]

    K. Chen, S. Xie, Z. Ma, P. R. Sanketi, and K. Goldberg. Robo2vlm: Visual question answering from large-scale in-the-wild robot manipulation datasets.arXiv preprint arXiv:2505.15517, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.Int. J. Robot. Res., 2023

    2023

  12. [12]

    C. Cui, P. Ding, W. Song, et al. OpenHelix: A Short Survey, Empirical Analysis, and Open- Source Dual-System VLA Model for Robotic Manipulation.arXiv preprint arXiv:2505.03912, 2025. 12

    Pith/arXiv arXiv 2025

  13. [13]

    Dalal, U

    D. Dalal, U. Mishra, N. Ahuja, and N. Jojic. City Navigation in the Wild: Exploring Emergent Navigation from Web-Scale Knowledge in MLLMs.arXiv preprint arXiv:2512.15933, 2025

    arXiv 2025

  14. [14]

    Dalal, G

    D. Dalal, G. Vashishtha, U. Mishra, J. Kim, M. Kanda, H. Ha, S. Lazebnik, H. Ji, and U. Jain. Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping.arXiv preprint arXiv:2510.09741, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Dalal, M

    D. Dalal, M. Kanda, Z. Wang, H. Ji, and U. Jain. Compositional Reasoning via Joint Im- age and Language Decomposition. In V . Demberg, K. Inui, and L. Marquez, editors,Find- ings of the Association for Computational Linguistics: EACL 2026, pages 5753–5775, Ra- bat, Morocco, Mar. 2026. Association for Computational Linguistics. ISBN 979-8-89176- 386-9. doi...

    work page doi:10.18653/v1/2026.findings-eacl.304 2026

  16. [16]

    A. Das, S. Datta, G. Gkioxari, S. Lee, D. Parikh, and D. Batra. Embodied question answering. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1–10, 2018

    2018

  17. [17]

    Dasari, M

    S. Dasari, M. K. Srirama, U. Jain, and A. Gupta. An Unbiased Look at Datasets for Visuo- Motor Pre-Training. InConference on Robot Learning, pages 1183–1198, 2023

    2023

  18. [18]

    Doshi, H

    R. Doshi, H. Walke, O. Mees, S. Dasari, and S. Levine. Scaling cross-embodied learn- ing: One policy for manipulation, navigation, locomotion and aviation.arXiv preprint arXiv:2408.11812, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Douillard, M

    A. Douillard, M. Cord, C. Ollion, T. Robert, and E. Valle. PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning. InEuropean Conference on Computer Vision, 2020

    2020

  20. [20]

    Driess, F

    D. Driess, F. Xia, M. S. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, et al. Palm-e: An embodied multimodal language model.arXiv preprint arXiv:2303.03378, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Driess, J

    D. Driess, J. T. Springenberg, B. Ichter, L. Yu, A. Li-Bell, K. Pertsch, A. Z. Ren, H. Walke, Q. Vuong, L. X. Shi, et al. Knowledge insulating vision-language-action models: Train fast, run fast, generalize better.arXiv preprint arXiv:2505.23705, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    S. Fei, S. Wang, J. Shi, Z. Dai, J. Cai, P. Qian, L. Ji, X. He, S. Zhang, Z. Fei, J. Fu, J. Gong, and X. Qiu. LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models. arXiv preprint arXiv:2510.13626, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Florence, C

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mor- datch, and J. Tompson. Implicit behavioral cloning. InConference on Robot Learning, pages 158–168. PMLR, 2022

    2022

  24. [24]

    Q. Gao, X. Pi, K. Liu, J. Chen, R. Yang, X. Huang, X. Fang, L. Sun, G. Kishore, B. Ai, et al. Do vision-language models have internal world models? towards an atomic evaluation.arXiv preprint arXiv:2506.21876, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Goyal, H

    A. Goyal, H. Hadfield, X. Yang, V . Blukis, and F. Ramos. VLA-0: Building State-of-the-Art VLAs with Zero Modification.arXiv preprint arXiv:2510.13054, 2025

    arXiv 2025

  26. [26]

    Gretton, O

    A. Gretton, O. Bousquet, A. Smola, and B. Sch ¨olkopf. Measuring Statistical Dependence with Hilbert–Schmidt Norms. InInt. Conf. Algorithmic Learn. Theory, pages 63–77. Springer, 2005

    2005

  27. [27]

    Grover, A

    S. Grover, A. Gopalkrishnan, B. Ai, H. I. Christensen, H. Su, and X. Li. Enhancing Gener- alization in Vision-Language-Action Models by Preserving Pretrained Representations.arXiv preprint arXiv:2509.11417, 2025. 13

    arXiv 2025

  28. [28]

    Hinton, O

    G. Hinton, O. Vinyals, and J. Dean. Distilling the Knowledge in a Neural Network.arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  29. [29]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-Rank Adaptation of Large Language Models. InICLR, 2022

    2022

  30. [30]

    Huang, M

    C. Huang, M. M. Zhang, R. Azarcon, G. Chou, and Z. Kira. MAPS: Preserving Vision-Language Representations via Module-Wise Proximity Scheduling for Better Vision- Language-Action Generalization.arXiv preprint arXiv:2511.19878, 2025

    arXiv 2025

  31. [31]

    D. A. Hudson and C. D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700–6709, 2019

    2019

  32. [32]

    J. Jang, S. Ye, Z. Lin, J. Xiang, J. Bjorck, Y . Fang, F. Hu, S. Huang, K. Kundalia, Y .-C. Lin, L. Magne, A. Mandlekar, A. Narayan, Y . L. Tan, G. Wang, J. Wang, Q. Wang, Y . Xu, X. Zeng, K. Zheng, R. Zheng, M.-Y . Liu, L. Zettlemoyer, D. Fox, J. Kautz, S. Reed, Y . Zhu, and L. Fan. DreamGen: Unlocking generalization in robot learning through video world...

    Pith/arXiv arXiv 2025

  33. [33]

    X. Jiao, Y . Yin, L. Shang, X. Jiang, X. Chen, L. Li, F. Wang, and Q. Liu. TinyBERT: Distilling BERT for Natural Language Understanding. InFindings of the Association for Computational Linguistics: EMNLP 2020, 2020

    2020

  34. [34]

    Kachaev, M

    N. Kachaev, M. Kolosov, D. Zelezetsky, A. K. Kovalev, and A. I. Panov. Don’t Blind Your VLA: Aligning Visual Representations for OOD Generalization.arXiv preprint arXiv:2510.25616, 2025

    arXiv 2025

  35. [35]

    Karamcheti, S

    S. Karamcheti, S. Nair, A. Balakrishna, P. Liang, T. Kollar, and D. Sadigh. Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models. InInternational Conference on Machine Learning, 2024

    2024

  36. [36]

    M. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Fos- ter, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. OpenVLA: An Open-Source Vision-Language-Action Model. InCon- ference on Robot Learning, 2024

    2024

  37. [37]

    M. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Kirkpatrick, R

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. Overcoming Catastrophic Forgetting in Neural Networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017

    2017

  39. [39]

    Kornblith, M

    S. Kornblith, M. Norouzi, H. Lee, and G. Hinton. Similarity of Neural Network Representa- tions Revisited. InInt. Conf. Mach. Learn., pages 3519–3529, 2019

    2019

  40. [40]

    J. Lee, J. Duan, H. Fang, Y . Deng, S. Liu, B. Li, B. Fang, J. Zhang, Y . R. Wang, S. Lee, et al. Molmoact: Action reasoning models that can reason in space.arXiv preprint arXiv:2508.07917, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Li and D

    Z. Li and D. Hoiem. Learning without Forgetting.IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935–2947, 2017

    2017

  42. [42]

    X. Li, Y . Grandvalet, and F. Davoine. Explicit Inductive Bias for Transfer Learning with Convolutional Networks. InInternational Conference on Machine Learning, 2018. 14

    2018

  43. [43]

    X. Li, M. Liu, H. Zhang, C. Yu, J. Xu, H. Wu, C. Cheang, Y . Jing, W. Zhang, H. Liu, et al. Vision-language foundation models as effective robot imitators.arXiv preprint arXiv:2311.01378, 2023

    Pith/arXiv arXiv 2023

  44. [44]

    X. Li, M. Liu, H. Zhang, C. Yu, J. Xu, H. Wu, C. Cheang, Y . Jing, W. Zhang, H. Liu, H. Li, and T. Kong. Vision-Language Foundation Models as Effective Robot Imitators. InICLR, 2024

    2024

  45. [45]

    S. Li, Y . Gao, D. Sadigh, and S. Song. Unified Video Action Model.arXiv preprint arXiv:2503.00200, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Liebel and M

    L. Liebel and M. K ¨orner. Auxiliary tasks in multi-task learning.arXiv preprint arXiv:1805.06334, 2018

    Pith/arXiv arXiv 2018

  47. [47]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll´ar, and C. L. Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014

    2014

  48. [48]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. InNeurIPS, 2023

    2023

  49. [49]

    F. Liu, G. Emerson, and N. Collier. Visual Spatial Reasoning.Transactions of the Association for Computational Linguistics, 11, 2023

    2023

  50. [50]

    Majumdar, A

    A. Majumdar, A. Ajay, X. Zhang, P. Putta, S. Yenamandra, M. Henaff, S. Silwal, P. Mcvay, O. Maksymets, S. Arnaud, et al. Openeqa: Embodied question answering in the era of foun- dation models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16488–16498, 2024

    2024

  51. [51]

    Mukhoti, Y

    J. Mukhoti, Y . Gal, P. H. S. Torr, and P. K. Dokania. Fine-tuning can cripple your foundation model; preserving features may be the solution.Transactions on Machine Learning Research, 2024

    2024

  52. [52]

    S. Nair, A. Rajeswaran, V . Kumar, C. Finn, and A. Gupta. R3M: A universal visual represen- tation for robot manipulation. InConference on Robot Learning, 2022

    2022

  53. [53]

    O’Neill, A

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

    2024

  54. [54]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haz- iza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024

    2024

  55. [55]

    Pertsch, K

    K. Pertsch, K. Stachowicz, B. Ichter, D. Driess, S. Nair, Q. Vuong, O. Mees, C. Finn, and S. Levine. FAST: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    D. Qu, H. Song, Q. Chen, Z. Chen, X. Gao, X. Ye, Q. Lv, M. Shi, G. Ren, C. Ruan, et al. EO- 1: An Open Unified Embodied Foundation Model for General Robot Control.arXiv preprint arXiv:2508.21112, 2025

    arXiv 2025

  57. [57]

    D. Qu, H. Song, Q. Chen, Y . Yao, X. Ye, Y . Ding, Z. Wang, J. Gu, B. Zhao, D. Wang, et al. Spatialvla: Exploring spatial representations for visual-language-action model.arXiv preprint arXiv:2501.15830, 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Radosavovic, B

    I. Radosavovic, B. Shi, L. Fu, K. Goldberg, T. Darrell, and J. Malik. Robot Learning with Sensorimotor Pre-training. InConference on Robot Learning, pages 683–693, 2023. 15

    2023

  59. [59]

    Radosavovic, T

    I. Radosavovic, T. Xiao, S. James, P. Abbeel, J. Malik, and T. Darrell. Real-World Robot Learning with Masked Visual Pre-training. InConference on Robot Learning, pages 416–426, 2023

    2023

  60. [60]

    Ramos, S

    S. Ramos, S. Girgin, L. Hussenot, D. Vincent, H. Yakubovich, D. Toyama, A. Gergely, P. Stanczyk, R. Marinier, J. Harmsen, et al. Rlds: an ecosystem to generate, share and use datasets in reinforcement learning.arXiv preprint arXiv:2111.02767, 2021

    Pith/arXiv arXiv 2021

  61. [61]

    S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-Maron, M. Gimenez, Y . Sulsky, J. Kay, J. T. Springenberg, et al. A generalist agent.arXiv preprint arXiv:2205.06175, 2022

    Pith/arXiv arXiv 2022

  62. [62]

    Reuss, J

    M. Reuss, J. Pari, P. Agrawal, and R. Lioutikov. Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning. InICLR, 2025

    2025

  63. [63]

    Rigter, T

    M. Rigter, T. Gupta, A. Hilmkil, and C. Ma. A VID: Adapting video diffusion models to world models. InInternational Conference on Learning Representations, 2025

    2025

  64. [64]

    Romero, N

    A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y . Bengio. FitNets: Hints for Thin Deep Nets. InInternational Conference on Learning Representations, 2015

    2015

  65. [65]

    Routray, H

    S. Routray, H. Pan, U. Jain, S. Bahl, and D. Pathak. ViPRA: Video prediction for robot actions. arXiv preprint arXiv:2511.07732, 2025

    arXiv 2025

  66. [66]

    Shukor, D

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, and A. Marafioti. SmolVLA: A vision-language-action model for afford- able and efficient robotics.arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  67. [67]

    Singh, V

    A. Singh, V . Natarajan, M. Shah, Y . Jiang, X. Chen, D. Batra, D. Parikh, and M. Rohrbach. Towards vqa models that can read. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317–8326, 2019

    2019

  68. [68]

    W. Song, Z. Zhou, H. Zhao, et al. ReconVLA: Reconstructive Vision-Language-Action Model as Effective Robot Perceiver.arXiv preprint arXiv:2508.10333, 2025

    Pith/arXiv arXiv 2025

  69. [69]

    StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing.arXiv preprint arXiv:2604.05014, 2026

    StarVLA Community. StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing.arXiv preprint arXiv:2604.05014, 2026

    Pith/arXiv arXiv 2026

  70. [70]

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  71. [71]

    L. Wang, X. Chen, J. Zhao, and K. He. Scaling Proprioceptive-Visual Learning with Hetero- geneous Pre-trained Transformers. InAdvances in Neural Information Processing Systems, 2024

    2024

  72. [72]

    Y . Wang, P. Ding, L. Li, C. Cui, Z. Ge, X. Tong, W. Song, H. Zhao, W. Zhao, P. Hou, et al. Vla- adapter: An effective paradigm for tiny-scale vision-language-action model.arXiv preprint arXiv:2509.09372, 2025

    arXiv 2025

  73. [73]

    J. Wen, Y . Zhu, J. Li, Z. Tang, C. Shen, and F. Feng. Dexvla: Vision-language model with plug-in diffusion expert for general robot control.arXiv preprint arXiv:2502.05855, 2025

    Pith/arXiv arXiv 2025

  74. [74]

    P. Wu, Y . Shentu, Z. Yi, X. Lin, and P. Abbeel. Gello: A general, low-cost, and intuitive teleoperation framework for robot manipulators. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12156–12163. IEEE, 2024. 16

    2024

  75. [75]

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, X. Ren, X. Ren, Y . Fan, Y . Su, Y . Zhang, Y . Wan, Y . Liu, Z. Cui, Z. Zhang, Z. Qiu, e...

    Pith/arXiv arXiv 2024

  76. [76]

    J. Yang, R. Tan, Q. Wu, R. Zheng, B. Peng, Y . Liang, Y . Gu, M. Cai, S. Ye, J. Jang, et al. Magma: A foundation model for multimodal ai agents. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14203–14214, 2025

    2025

  77. [77]

    Y . Yue, Y . Wang, B. Kang, Y . Han, S. Wang, S. Song, J. Feng, and G. Huang. DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Execution. In NeurIPS, 2024

    2024

  78. [78]

    Zawalski, W

    M. Zawalski, W. Chen, K. Pertsch, O. Mees, C. Finn, and S. Levine. Robotic control via embodied chain-of-thought reasoning.arXiv preprint arXiv:2407.08693, 2024

    Pith/arXiv arXiv 2024

  79. [79]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre- training. InProceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023

    2023

  80. [80]

    X. Zhou, Y . Xu, G. Tie, Y . Chen, G. Zhang, D. Chu, P. Zhou, and L. Sun. LIBERO-PRO: To- wards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization. arXiv preprint arXiv:2510.03827, 2025

    Pith/arXiv arXiv 2025

Showing first 80 references.