The reviewed record of science sign in
Pith

arxiv: 2607.06403 · v1 · pith:MVL6R7YR · submitted 2026-07-07 · cs.RO

From Foundation to Application: Improving VLA Models in Practice

Wei Wu , Fangjing Wang , Fan Lu , He Sun , Shi Liu , Yunnan Wang , Yibin Yan , Yong Wang
show 16 more authors
Shuailei Ma Xinyang Wang Yibin Liu Shuai Yang Tianxiang Zhou Kejia Zhang Lei Zhou Cheng Su Nan Xue Bin Tan Han Zhang Youchao Zhang Fei Liao Xing Zhu Yujun Shen Kecheng Zheng
This is my paper

Reviewed by Pith2026-07-08 06:48 UTCglm-5.2pith:MVL6R7YRopen to challenge →

classification cs.RO
keywords vision-language-actionrobot manipulationmixture of expertsdual-query distillationcross-embodiment generalizationmobile manipulationwhole-body controlVLA foundation model
0
0 comments X

The pith

Robot system handles whole-body tasks across 20 platforms

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

The paper presents LingBot-VLA 2.0, a vision-language-action (VLA) model for robotic control that argues practical deployment requires three coordinated improvements over lab-grade systems: broader multi-embodiment pretraining data (60,000 hours spanning 20 robot configurations and egocentric human video), an expanded action space covering whole-body degrees of freedom (head, waist, mobile base, dexterous hands beyond standard dual-arm), and predictive dynamics modeling via a dual-query distillation scheme where a depth teacher provides geometric supervision and a video teacher provides temporal supervision. The central claim is that these three modifications jointly bridge the lab-to-deployment gap. The paper evaluates on the GM-100 bimanual benchmark and two long-horizon mobile manipulation tasks, reporting best-overall performance against π0.5 and the prior LingBot-VLA-1.0, with gains most pronounced on tasks requiring object grounding and future-aware planning. The architecture uses token-level sparse mixture-of-experts layers in the action expert with auxiliary-loss-free load balancing, and the dual-query distillation appends learnable current and future queries to the VLM's token sequence, training them to predict depth and video features at both timesteps.

Core claim

The paper's central object is a VLA system that unifies three deployment-oriented modifications: a 55-dimensional canonical action vector mapping heterogeneous embodiment controls (arms, end-effectors, grippers, hands, waist, head, mobility) into a shared representation; a token-level sparse MoE action expert with sigmoid routing and bias-based load balancing that decouples expert assignment from expert weighting; and a dual-query distillation framework where current and future query tokens are supervised by a depth model (LingBot-Depth) for geometric structure and a causal video model (DINO-Video, built on DINOv3 with causal temporal attention and 3D rotary embeddings) for temporal dynamics

What carries the argument

LingBot-VLA 2.0

Where Pith is reading between the lines

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

  • The paper does not isolate the three headline contributions (data scale, action space expansion, predictive dynamics) against each other in a controlled ablation, so the relative contribution of each remains unclear from the evidence presented.
  • The comparison with π0.5 is not controlled for training data volume (60k hours vs. undisclosed), so gains could partly reflect data scale rather than architectural choices.
  • The action-space ablations (relative vs absolute, EEF vs joint, normalization, loss) examine micro-level representation choices on four tasks but do not test the three headline modifications.
  • If the predictive dynamics mechanism generalizes, it could be tested as a plug-in module on other VLA architectures to see whether future-query distillation consistently improves contact-rich or long-horizon tasks.

Load-bearing premise

The paper assumes the reported performance gains stem from its three proposed modifications working in concert, but it never runs an ablation isolating data scale, action-space expansion, and predictive dynamics against data-matched baselines.

What would settle it

If a baseline VLA trained on a comparable volume of multi-embodiment data with standard dual-arm action spaces but without the dual-query distillation or MoE architecture achieved similar GM-100 and long-horizon scores, the three headline contributions would lose their causal support.

Figures

Figures reproduced from arXiv: 2607.06403 by Bin Tan, Cheng Su, Fangjing Wang, Fan Lu, Fei Liao, Han Zhang, He Sun, Kecheng Zheng, Kejia Zhang, Lei Zhou, Nan Xue, Shi Liu, Shuailei Ma, Shuai Yang, Tianxiang Zhou, Wei Wu, Xing Zhu, Xinyang Wang, Yibin Liu, Yibin Yan, Yong Wang, Youchao Zhang, Yujun Shen, Yunnan Wang.

Figure 1
Figure 1. Figure 1: Overview of LingBot-VLA 2.0. We revamp the data processing pipeline and curate 60,000 hours of pretraining data, including 50,000 hours of robot trajectories across 20 robot configurations and 10,000 hours of egocentric human videos. Moreover, our model supports degrees of freedom for the head, waist, mobile base, and dexterous hands, enabling robots to handle more complex real-world tasks. We formulate fu… view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the pre-training dataset used by LingBot-VLA-2.0. The dataset includes 20 robot embodiments with degrees of freedom in the arms, heads, waists, mobile bases, and dexterous hands. 2 Related Work 2.1 Generalist Robot Manipulation Policies Recent vision-language-action (VLA) models [4–8, 10, 14, 30, 32–34, 39, 41, 42, 47] have established foundation VLA model as a promising paradigm for gener… view at source ↗
Figure 3
Figure 3. Figure 3: Data processing pipeline. 2.2 Mixture-of-Experts Vision-Language-Action Model Mixture-of-Experts (MoE) has emerged as a pivotal mechanism for scaling action modeling capacity in VLA systems. For contact-rich manipulation, ForceVLA [45] and ForceVLA2 [17] introduce force-aware MoE modules to fuse sparse but critical force feedback with visual-language representations, while MoDE-VLA [29] further exploits fo… view at source ↗
Figure 4
Figure 4. Figure 4: Unified action representation. We map heterogeneous embodiment controls into compact action vectors composed of shared body-part components. estimation. After VLM pre-filtering, we process the remaining data according to whether action or hand trajectory labels are available. For data with existing action or hand trajectory labels, including action-labeled open-source datasets and in-house data, we perform… view at source ↗
Figure 5
Figure 5. Figure 5: Per-action statistics of the subtask annotations: duration composition, frequency (fraction of all subtasks), total time (fraction of the total annotated time), and mean duration. Actions are grouped as in Tab. 2; idle is filtered out of the training data and omitted here. cup bottle box bowl water bottle can bag block cylinder drawer plate ball marker spoon tool case cap pencube plush toy lid packet screw… view at source ↗
Figure 6
Figure 6. Figure 6: Word cloud of the manipulated objects in the subtask annotations, sized by frequency. Given the modulated FFN input uℓ,t ∈ R d of token t at layer ℓ, the MoE layer computes mℓ(uℓ,t) = E (s) ℓ (uℓ,t) + λ X j∈R(uℓ,t) gℓ,j (uℓ,t)E (r) ℓ,j (uℓ,t), (2) where E (s) ℓ (·) denotes the shared expert, E (r) ℓ,j (·) denotes the j-th routed expert, R(uℓ,t) is the selected top-K routed expert set, and λ is a routed-out… view at source ↗
Figure 7
Figure 7. Figure 7: Comparable active-parameter scaling comparison between Dense model and MoE model using training loss on pre-training data and validation action error on GM-100 tasks. 4.2 Spatiotemporal-Aware VLA via Dual-Query Distillation To promote both geometric awareness and causal temporal understanding in the LingBot-VLA 2.0, we adopt a dual￾query distillation framework inspired by recent works [35], where LingBot-D… view at source ↗
Figure 8
Figure 8. Figure 8: Per-subtask performance on the long-horizon mobile manipulation benchmark. Bars report the mean subtask completion score (%) across 15 trials under both in-domain (ID) and out-of-distribution (OOD) settings. score and success rate, indicating that the model often makes partial progress but fails at the final precise placement, release, or completion step. Moreover, the performance disparity between Agilex … view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of the long-horizon mobile manipulation experiments platform. 6 Ablation Studies 6.1 Action Space Action target. For the action target, relative joint actions substantially outperform absolute joint actions, improving the average success rate from 33.7 to 55.0. This is supported by the action statistics in Fig. 12A. Across the four tasks, the standard deviation of relQpos is only 31%–37% of th… view at source ↗
Figure 10
Figure 10. Figure 10: Results on the four GM-100 real-robot tasks. L1 L2 L3 L4 L5 L6 L7 R1 R2 R3 R4 R5 R6 R7 −3 −2 −1 0 1 2 3 Joint relQpos MeanStd-normalized value Pooled ALL N=360k Pooled ALL L1 L2 L3 L4 L5 L6 L7 R1 R2 R3 R4 R5 R6 R7 SR 58.7 Gap 0.68 Barcode Scan L1 L2 L3 L4 L5 L6 L7 R1 R2 R3 R4 R5 R6 R7 SR 42.7 Gap 0.55 Scoop Rice L1 L2 L3 L4 L5 L6 L7 R1 R2 R3 R4 R5 R6 R7 SR 41.7 Gap 1.59 Squeeze Ketchup L1 L2 L3 L4 L5 L6 L… view at source ↗
Figure 11
Figure 11. Figure 11: Per-dimension MeanStd-normalized action distributions for joint and EEF action spaces. The first column shows the pooled distribution over all four tasks, while the remaining columns show task-specific distributions. The gap measures task-to-pooled distribution alignment based on per-dimension median and IQR differences. Action space. For the action space, EEF and joint actions obtain similar average succ… view at source ↗
Figure 12
Figure 12. Figure 12: Action targets and normalization statistics for the four GM-100 tasks. Left: relQpos substantially reduces the action scale compared with absQpos. Right: normalization changes the effective dynamic range of MeanStd-normalized relQpos; Q01–Q99 is shown in the actual unclipped setting. Normalization. For normalization, Figure 12B shows the normalized relQpos distributions under the three schemes. MinMax com… view at source ↗
Figure 13
Figure 13. Figure 13: Causal perception results of LingBot-VLA 2.0 under visual distillation. Left: depth/DINO-Video-PCA ground truth and prediction for current image. Right: depth/DINO-Video-PCA ground truth and prediction for future image. 7 Conclusion LingBot-VLA 2.0 narrows the gap between VLA foundation models and real-world robotic deployment by improving generalization, expanding whole-body action modeling, and strength… view at source ↗
read the original abstract

Despite recent progress of VLA foundation models, the disparity between laboratory conditions and real-world applications continues to impede their practical implementation. To bridge this gap, we present LingBot-VLA 2.0, which advances LingBot-VLA through improvements in three functional domains. (1) Generalization across tasks and embodiments. Compared to the previous version, we revamp the data processing pipeline and curate around 60,000 hours of data for pretraining, including 50,000 hours of robot trajectories spanning 20 robot configurations and 10,000 hours of egocentric human videos. (2) Expanded action space in addition to dual-arm hardware platforms. In particular, our system accommodates degrees of freedom for the heads, waists, mobile bases, and dexterous hands, thereby empowering the robots to tackle more complex tasks in practical scenarios. (3) Predictive dynamics modeling for improved temporal reasoning. Specifically, we formulate future prediction as a proxy task, facilitated by a video representation model for semantic priors and a depth estimation model for geometric cues. Evaluations on the GM-100 benchmark, conducted in a generalist setting, validate the beneficial impact of these proposed modifications. Furthermore, benefiting from the expanded pretraining data that covers whole-body degrees of freedom, LingBot-VLA-2.0 demonstrates strong cross-embodiment long-horizon mobile manipulation capability across the two robotic platforms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

4 major / 8 minor

Summary. The paper presents LingBot-VLA 2.0, an improved vision-language-action (VLA) model targeting the gap between laboratory benchmarks and real-world robotic deployment. The system advances three areas: (1) a curated 60,000-hour multi-embodiment pretraining dataset (50k robot + 10k egocentric), (2) an expanded 55-dimensional action space supporting whole-body DoF (head, waist, mobile base, dexterous hands), and (3) a dual-query distillation framework using LingBot-Depth and DINO-Video as teachers for predictive dynamics modeling. The architecture uses token-level sparse MoE layers in the action expert with auxiliary-loss-free load balancing. Evaluation is conducted on nine GM-100 bimanual tasks under a generalist setting on two platforms (Agilex Cobot Magic, Galaxea R1 Pro), plus two long-horizon mobile manipulation tasks. The authors report improvements over LingBot-VLA-1.0 and π0.5 on aggregate metrics.

Significance. The paper makes a substantial engineering contribution to the VLA field. The 60k-hour curated dataset across 20 embodiments is a significant resource, and the unified 55-D action representation is a clean design for heterogeneous platforms. The MoE scaling experiment (Fig. 7) with matched active parameters provides a fair comparison. The dual-query distillation framework is well-motivated, and the DINO-Video teacher is benchmarked on LARYBench (Tab. 3). The release of checkpoints and code is commendable and supports reproducibility. The long-horizon mobile manipulation results with 15-trial evaluation and ID/OOD settings add practical value. However, the significance of the three headline contributions is undermined by the absence of contribution-level ablations, as detailed below.

major comments (4)
  1. §5.2, Table 5: The central claim that three coordinated modifications (data scale, expanded action space, predictive dynamics) jointly improve performance is not supported by a contribution-level ablation. The ablations in §6.1 only examine micro-level action representation choices (relative vs absolute, EEF vs joint, normalization, loss) on four tasks. No baseline → +data → +action-space → +predictive-dynamics staircase is run. Without isolating these three contributions, the causal attribution to all three modifications working in concert is unsupported. This is the paper's central thesis.
  2. §5.2, Table 5: The headline comparison with π0.5 is not controlled for training data volume. LingBot-VLA-2.0 uses 60k hours; π0.5's data volume is undisclosed. The aggregate improvements (66.2/34.4 vs 59.1/32.2 on Agilex) could largely reflect data scale rather than architectural innovation. A data-matched comparison or at minimum explicit discussion of this confound is needed to support the claim that the three modifications are responsible for the gains.
  3. §5.2, Table 5: Per-task variance is high and the aggregate average may be misleading. On Agilex, LingBot-VLA-2.0 dramatically beats π0.5 on Retrieve keychain (100/100 vs 20/20) but loses on Block sorting (56.8/0.0 vs 90.4/60.0), Sort snacks (66.2/10.0 vs 82.4/30.0), Pack eggs (44.4/0.0 vs 72.4/20.0), and Tool packing (60/0 vs 66/30). No error bars or trial counts are reported for GM-100 (only long-horizon tasks specify 15 trials). The statistical robustness of the headline comparison is unclear without these details.
  4. §4.2, Eqs. (9)–(10): The dual-query distillation uses LingBot-Depth [28] and DINO-Video as teachers, both of which are co-authored by the same team. While this is disclosed, the paper does not provide an ablation isolating the contribution of predictive dynamics distillation to the final action performance. Fig. 13 shows perceptual quality of the distilled queries, but no experiment links this to improved manipulation success. A simple ablation (with vs without distillation) on a few GM-100 tasks would substantially strengthen the claim.
minor comments (8)
  1. §3.1.1: The data filtering thresholds (jerk Z-score, static proportion >95%, valid-frame ratio <20%) are described qualitatively but not specified numerically. Stating the actual threshold values would aid reproducibility.
  2. §4.1: The MoE hyperparameters (K, Nr, λ, γ) are mentioned in the equations but their specific values for the final model are not stated in the text. Please specify.
  3. Table 1: The 'Body DoF' column includes arm DoF in some rows but not others, and the relationship between 'Body DoF' and 'Total DoF' is not always consistent (e.g., AgiBot G1: Body DoF 4, Total 20, but Arm DoF 14 + Body 4 = 18, not 20). Clarifying the definitions would help.
  4. §3.1.3: The 55-D action vector includes 4 reserved dimensions. It would help to state what these are reserved for.
  5. Table 5: The 'Overall average' row appears to average 9 tasks, but 10 task rows are listed for Agilex (BM-19 through BM-107). Please clarify which tasks are averaged and why one is excluded.
  6. §5.3, Table 6: The long-horizon comparison only includes π0.5, not LingBot-VLA-1.0. Adding the v1.0 comparison or explaining its absence would strengthen the comparison.
  7. Fig. 7: The x-axis label 'Training step (k)' is ambiguous — are these optimizer steps or epochs? Specifying would help interpretation.
  8. The paper mentions 'Qwen3.6-27B' [25] for annotation but the reference appears to be a future-dated preprint. Confirming the correct model version and citation would be appropriate.

Simulated Author's Rebuttal

4 responses · 2 unresolved

We thank the referee for a thorough and constructive report. The referee acknowledges the engineering contributions (dataset, unified action space, MoE scaling, dual-query distillation, reproducible release) but identifies a central concern: the absence of contribution-level ablations isolating the three headline modifications (data scale, expanded action space, predictive dynamics). Additional concerns address the uncontrolled data-volume confound in the π0.5 comparison, missing trial counts and error bars for GM-100, and the lack of a distillation ablation linking perceptual quality to manipulation success. We agree that contribution-level ablations and statistical reporting are needed and will add them. On the data-volume confound, we provide honest discussion of what can and cannot be controlled given π0.5's undisclosed training data.

read point-by-point responses

  1. Referee: §5.2, Table 5: The central claim that three coordinated modifications (data scale, expanded action space, predictive dynamics) jointly improve performance is not supported by a contribution-level ablation. The ablations in §6.1 only examine micro-level action representation choices (relative vs absolute, EEF vs joint, normalization, loss) on four tasks. No baseline → +data → +action-space → +predictive-dynamics staircase is run. Without isolating these three contributions, the causal attribution to all three modifications working in concert is unsupported. This is the paper's central thesis.

    Authors: The referee is correct that the current manuscript does not include a contribution-level staircase ablation isolating the three headline modifications. The ablations in §6.1 address micro-level action representation choices, which are orthogonal to the three functional-domain contributions. We agree this is a gap relative to the paper's central thesis. In the revision, we will add a contribution-level ablation on a representative subset of GM-100 tasks, incrementally adding: (1) expanded pretraining data (from LingBot-VLA-1.0's data to the 60k-hour corpus), (2) the expanded 55-D action space, and (3) dual-query predictive dynamics distillation. We will report both progress score and success rate at each stage. We acknowledge that a fully controlled staircase on all nine GM-100 tasks across both platforms may not be feasible within the revision timeline due to compute and robot evaluation constraints, so we will be explicit about the subset and platform used. We will also temper the language in the abstract and conclusion to accurately reflect what the ablation does and does not prove about the joint contribution of all three modifications. revision: yes

  2. Referee: §5.2, Table 5: The headline comparison with π0.5 is not controlled for training data volume. LingBot-VLA-2.0 uses 60k hours; π0.5's data volume is undisclosed. The aggregate improvements (66.2/34.4 vs 59.1/32.2 on Agilex) could largely reflect data scale rather than architectural innovation. A data-matched comparison or at minimum explicit discussion of this confound is needed to support the claim that the three modifications are responsible for the gains.

    Authors: This is a fair concern. π0.5's training data volume and composition are not publicly disclosed, so we cannot run a data-matched comparison against π0.5 directly. We will add an explicit discussion of this confound in the revision, acknowledging that the aggregate improvements over π0.5 cannot be cleanly attributed to architectural innovation versus data scale. To partially address this, the contribution-level ablation described in our response to the first comment will include a LingBot-VLA-1.0 → +data-only condition, which isolates the effect of scaling from 1.0's data to 60k hours without the other two modifications. This provides an internal data-scale control, even if an external data-matched comparison with π0.5 is not possible. We will also revise the language in §5.2 to avoid implying that the gains over π0.5 are solely due to the three architectural modifications. revision: partial

  3. Referee: §5.2, Table 5: Per-task variance is high and the aggregate average may be misleading. On Agilex, LingBot-VLA-2.0 dramatically beats π0.5 on Retrieve keychain (100/100 vs 20/20) but loses on Block sorting (56.8/0.0 vs 90.4/60.0), Sort snacks (66.2/10.0 vs 82.4/30.0), Pack eggs (44.4/0.0 vs 72.4/20.0), and Tool packing (60/0 vs 66/30). No error bars or trial counts are reported for GM-100 (only long-horizon tasks specify 15 trials). The statistical robustness of the headline comparison is unclear without these details.

    Authors: The referee is correct that GM-100 trial counts and error bars are not reported in the current manuscript, and that per-task variance is substantial. We will add trial counts for all GM-100 evaluations in the revised Table 5. For error bars, we will report standard deviations or confidence intervals where the evaluation protocol supports multiple independent trials. We acknowledge that some GM-100 tasks may have been evaluated with a limited number of trials due to robot availability and time constraints; where this is the case, we will state the trial count explicitly and note the resulting statistical limitations. We will also add a brief discussion of the per-task variance pattern the referee identifies, including the observation that LingBot-VLA-2.0 underperforms π0.5 on several tasks (Block sorting, Sort snacks, Pack eggs, Tool packing on Agilex), to provide a more balanced presentation rather than relying solely on aggregate averages. revision: yes

  4. Referee: §4.2, Eqs. (9)–(10): The dual-query distillation uses LingBot-Depth [28] and DINO-Video as teachers, both of which are co-authored by the same team. While this is disclosed, the paper does not provide an ablation isolating the contribution of predictive dynamics distillation to the final action performance. Fig. 13 shows perceptual quality of the distilled queries, but no experiment links this to improved manipulation success. A simple ablation (with vs without distillation) on a few GM-100 tasks would substantially strengthen the claim.

    Authors: We agree that linking the perceptual quality of distilled queries to downstream manipulation success is important and currently missing. The contribution-level ablation we propose in response to the first comment will include a with-distillation vs without-distillation comparison on a subset of GM-100 tasks, which directly addresses this concern. We will report both manipulation metrics (progress score, success rate) and, where feasible, perceptual metrics for the distilled queries. Regarding the use of co-authored teacher models: we chose LingBot-Depth and DINO-Video because they are specifically designed for robotics-relevant geometric and temporal supervision, and DINO-Video is benchmarked on LARYBench (Table 3) to demonstrate its standalone quality. However, we acknowledge that using independent teacher models would strengthen the claim that the distillation framework generalizes beyond our own models. In the revision, we will add a discussion of this limitation and, if compute permits, include a comparison using a non-co-authored depth teacher (e.g., Depth Anything) as an additional ablation point. revision: partial

standing simulated objections not resolved
  • The data-volume confound with π0.5 cannot be fully resolved because π0.5's training data volume and composition are not publicly disclosed. We can only provide internal data-scale controls (via the LingBot-VLA-1.0 → +data-only ablation) and explicit discussion of the limitation; a true data-matched external comparison is not possible without cooperation from the π0.5 authors.
  • A fully controlled contribution-level staircase ablation on all nine GM-100 tasks across both platforms may not be feasible within the revision timeline due to compute and robot evaluation constraints. We will conduct the ablation on a representative subset and be transparent about this scope limitation.

Circularity Check

0 steps flagged

No significant circularity; self-cited teacher models are used as tools, not as circular derivations

full rationale

The paper's three headline contributions (data pipeline, expanded action space, predictive dynamics via dual-query distillation) are validated on external benchmarks (GM-100, long-horizon mobile manipulation tasks). The self-cited components — LingBot-Depth [28] (co-authors overlap), DINO-Video (trained by authors), and the dual-query distillation framework [35] (co-author Yunnan Wang) — serve as teacher models and architectural scaffolding, not as the basis for a derivation that reduces to its own inputs. The distillation losses (Eqs. 9–10) are standard knowledge-distillation objectives: the student queries are trained to predict teacher outputs, which is explicitly framed as a proxy task, not as a novel first-principles prediction. DINO-Video is evaluated on the external LARYBench benchmark (Table 3), providing independent evidence of its quality as a teacher. The MoE vs Dense comparison (Fig. 7) is a genuine ablation with matched active parameters. The §6 ablations on action representation choices are independent experiments. While the absence of contribution-level ablations isolating the three headline modifications is a legitimate correctness concern (the causal attribution is unsupported), that is an ablation gap, not circularity. No step in the paper's chain reduces by construction to a self-cited input or a fitted parameter renamed as a prediction.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The axiom ledger captures the key assumptions and design choices that the paper relies on without independent validation.

free parameters (7)
  • MoE routing: K (top-K experts) = not specified numerically
    Number of routed experts activated per token; controls sparsity but exact value not stated in the paper
  • MoE routing: Nr (number of routed experts) = not specified numerically
    Total number of routed experts per MoE layer; not stated
  • λ (routed-output scaling factor) = not specified
    Scales the contribution of routed experts relative to shared expert (Eq. 2); value not reported
  • γ (bias update speed) = not specified
    Controls load-balancing bias update rate (Eq. 8); value not reported
  • 55-D action vector dimensionality allocation = 14 arm + 14 EEF + 2 gripper + 12 hand + 4 waist + 2 head + 3 mobility + 4 reserved
    Hand-designed to accommodate maximum dimensionality across embodiments; not derived from optimization
  • Action chunk size T (future prediction horizon) = not specified
    Determines the future query target in dual-query distillation; value not reported
  • Data filtering thresholds (jerk Z-score, static proportion, valid-frame ratio) = Z-score threshold (unspecified), 95% static, 20% valid frames
    Set per-embodiment for jerk/velocity/acceleration; exact Z-score threshold not reported
axioms (5)
  • domain assumption The 55-dimensional canonical action vector is a sufficient unified representation for all 20 embodiments
    §3.1.3, Fig. 4: The representation is hand-designed to accommodate maximum dimensionality. No proof or empirical validation that this particular allocation is optimal or sufficient is provided.
  • domain assumption Depth and video features from teacher models provide useful supervisory signal for action prediction
    §4.2: The dual-query distillation assumes that predicting future depth and DINO-Video features improves action quality. This is plausible but the paper does not ablate the distillation contribution against the other two headline improvements.
  • domain assumption Egocentric human hand trajectories, when transformed to camera coordinates, provide transferable action priors for robot control
    §3.1.2: The paper assumes that reconstructed MANO hand trajectories lifted to world coordinates and then transformed to camera coordinates (Eq. 1) are useful training signal for robot policies. The retargeting gap between human hands and robot grippers/hands is not explicitly addressed.
  • domain assumption Qwen3.6-27B produces sufficiently accurate subtask segmentation and instruction annotations for VLA pretraining
    §3.2: Automated annotation pipeline relies entirely on VLM-generated labels. No human accuracy audit or inter-annotator agreement is reported.
  • domain assumption DeepSeek-V3's auxiliary-loss-free load balancing mechanism transfers effectively from language to action token routing
    §4.1: The MoE design is directly adapted from DeepSeek-V3 [19]. The paper provides Fig. 7 showing lower training loss but does not analyze whether the load-balancing dynamics behave similarly for action tokens vs language tokens.
invented entities (2)
  • DINO-Video independent evidence
    purpose: Causal video representation model serving as temporal teacher for dual-query distillation
    Evaluated on LARYBench [24] (Tab. 3), an external benchmark, achieving best performance on 3 of 4 metrics. However, LARYBench itself may have overlap with the authors' evaluation ecosystem. Trained on 5M video clips from Internet, egocentric, and robotic data.
  • LingBot-Depth no independent evidence
    purpose: Depth estimation model serving as geometric teacher for dual-query distillation
    Referenced as [28] with a URL but no arXiv preprint or external benchmark evaluation is provided in the paper. The model is from the same author group. No independent validation of its quality is presented.

pith-pipeline@v1.1.0-glm · 23715 in / 5025 out tokens · 393311 ms · 2026-07-08T06:48:01.650351+00:00 · methodology

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

49 extracted references · 49 canonical work pages · 20 internal anchors

  1. [1]

    Cosmos 3: Omnimodal World Models for Physical AI

    Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Adeline Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yogesh Balaji, Josh Bapst, et al. Cosmos 3: Omnimodal world models for physical ai.arXiv preprint arXiv:2606.02800, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    Mahmoud Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Francois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khalidov, Patrick Labatut, Francisco Massa, Marc Szafraniec, Kapil Krishnakumar, Yong Li, X...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    HEX: Humanoid-Aligned Experts for Cross-Embodiment Whole-Body Manipulation

    Shuanghao Bai, Meng Li, Xinyuan Lv, Jiawei Wang, Xinhua Wang, Fei Liao, Chengkai Hou, Langzhe Gu, Wanqi Zhou, Kun Wu, Ziluo Ding, Zhiyuan Xu, Lei Sun, Shanghang Zhang, Zhengping Che, Jian Tang, and Badong Chen. HEX: Humanoid-aligned experts for cross-embodiment whole-body manipulation.arXiv preprint arXiv:2604.07993, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  4. [4]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. GR00T N1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Robert Equi, Chelsea Finn, Niccolo Fusai, Manuel Y . Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, brian ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Allen Z. Ren,...

    work page 2025

  6. [6]

    InProceedings of Robotics: Science and Systems, 2025

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.π0: A visio...

    work page 2025

  7. [7]

    Xiaomi-robotics-0: An open-sourced vision-language-action model with real-time execution.arXiv preprint arXiv:2602.12684, 2026

    Rui Cai, Jun Guo, Xinze He, Piaopiao Jin, Jie Li, Bingxuan Lin, Futeng Liu, Wei Liu, Fei Ma, Kun Ma, Feng Qiu, Heng Qu, Yifei Su, Qiao Sun, Dong Wang, Donghao Wang, Yunhong Wang, Rujie Wu, Diyun Xiang, Yu Yang, Hangjun Ye, Yuan Zhang, and Quanyun Zhou. Xiaomi-robotics-0: An open-sourced vision-language-action model with real-time execution.arXiv preprint ...

    work page arXiv 2026

  8. [8]

    GR-3 Technical Report

    Chilam Cheang, Sijin Chen, Zhongren Cui, Yingdong Hu, Liqun Huang, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Xiao Ma, et al. GR-3 technical report.arXiv preprint arXiv:2507.15493, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Lawam: Latent world action models for efficient dynamics-aware robot policies.arXiv preprint arXiv:2606.15768, 2026

    Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, Wenhao Tang, Zhiyuan Li, Weilin Liu, Zhuyu Yao, Boxun Li, et al. Lawam: Latent world action models for efficient dynamics-aware robot policies.arXiv preprint arXiv:2606.15768, 2026

    work page arXiv 2026

  10. [10]

    ABot-M0.5: Unified Mobility-and-Manipulation World Action Model

    Ronghan Chen, Yandan Yang, Zuojin Tang, Dongjie Huo, Tong Lin, Haoning Wu, Haoyun Liu, Yuzhi Chen, Lulu Zheng, Botai Yuan, Tianlun Li, Mingxin Wang, Dekang Qi, Bin Hu, Wei Mei, Yuze Xuan, Haolong Yang, Yanqing Zhu, Mu Xu, Zhiheng Ma, and Xinyuan Chang. Abot-m0.5: Unified mobility-and-manipulation world action model.arXiv preprint arXiv:2607.00678, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  11. [11]

    Dexworldmodel: Causal latent world modeling towards automated learning of embodied tasks, 2026

    DexForce AI Team of Physical AI. Dexworldmodel: Causal latent world modeling towards automated learning of embodied tasks, 2026. Technical report

    work page 2026

  12. [12]

    HiMoE-VLA: Hierarchical mixture-of-experts for generalist vision-language-action policies.arXiv preprint arXiv:2512.05693, 2025

    Zhiying Du, Bei Liu, Yaobo Liang, Yichao Shen, Haidong Cao, Xiangyu Zheng, Zhiyuan Feng, Zuxuan Wu, Jiaolong Yang, and Yu-Gang Jiang. HiMoE-VLA: Hierarchical mixture-of-experts for generalist vision-language-action policies.arXiv preprint arXiv:2512.05693, 2025

    work page arXiv 2025

  13. [13]

    Galaxea g0.5 technical report, 2026

    Galaxea Team. Galaxea g0.5 technical report, 2026. Technical report

    work page 2026

  14. [14]

    Galaxea Open-World Dataset and G0 Dual-System VLA Model

    Tao Jiang, Tianyuan Yuan, Yicheng Liu, Chenhao Lu, Jianning Cui, Xiao Liu, Shuiqi Cheng, Jiyang Gao, Huazhe Xu, and Hang Zhao. Galaxea open-world dataset and G0 dual-system vla model.arXiv preprint arXiv:2509.00576, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    RLDX-1 Technical Report

    Dongyoung Kim, Huiwon Jang, Myungkyu Koo, Suhyeok Jang, Taeyoung Kim, Beomjun Kim, Byungjun Yoon, Changsung Jang, Daewon Choi, Dongsu Han, Donguk Lee, Heeseung Kwon, Hojin Jeon, Jaehyun Kang, Jaekyoung Bae, Jihyuk Lee, Jimin Lee, John Won, Joonwoo Ahn, Junhyeong Park, Junyoung Sung, Kyungmin Lee, Minseong Han, Minsung Yoon, Sejune Joo, Seonil Son, Seungch...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    OpenVLA: An open-source vision-language-action model

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. OpenVLA: An open-source vision-language-action model. InConference on Robot Learning, 2025

    work page 2025

  17. [17]

    ForceVLA2: Unleashing hybrid force-position control with force awareness for contact-rich manipulation.arXiv preprint arXiv:2603.15169, 2026

    Yang Li, Zhaxizhuoma, Hongru Jiang, Junjie Xia, Hongquan Zhang, Jinda Du, Yunsong Zhou, Jia Zeng, Ce Hao, Jieji Ren, Qiaojun Yu, Cewu Lu, Yu Qiao, and Jiangmiao Pang. ForceVLA2: Unleashing hybrid force-position control with force awareness for contact-rich manipulation.arXiv preprint arXiv:2603.15169, 2026

    work page arXiv 2026

  18. [18]

    Holobrain-0 technical report.arXiv preprint arXiv:2602.12062, 2026

    Xuewu Lin, Tianwei Lin, Yun Du, Hongyu Xie, Yiwei Jin, Jiawei Li, Shijie Wu, Qingze Wang, Mengdi Li, Mengao Zhao, Ziang Li, Chaodong Huang, Hongzhe Bi, Lichao Huang, and Zhizhong Su. Holobrain-0 technical report.arXiv preprint arXiv:2602.12062, 2026

    work page arXiv 2026

  19. [19]

    DeepSeek-V3 Technical Report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos

    Hao Luo, Yicheng Feng, Wanpeng Zhang, Sipeng Zheng, Ye Wang, Haoqi Yuan, Jiazheng Liu, Chaoyi Xu, Qin Jin, and Zongqing Lu. Being-h0: Vision-language-action pretraining from large-scale human videos.arXiv preprint arXiv:2507.15597, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Being-h0.5: Scaling human-centric robot learning for cross-embodiment generalization

    Hao Luo, Ye Wang, Wanpeng Zhang, Sipeng Zheng, Ziheng Xi, Chaoyi Xu, Haiweng Xu, Haoqi Yuan, Chi Zhang, Yiqing Wang, Yicheng Feng, and Zongqing Lu. Being-h0.5: Scaling human-centric robot learning for cross-embodiment generalization. arXiv preprint arXiv:2601.12993, 2026

    work page arXiv 2026

  22. [22]

    Being-H0.7: A Latent World-Action Model from Egocentric Videos

    Hao Luo, Wanpeng Zhang, Yicheng Feng, Sipeng Zheng, Haiweng Xu, Chaoyi Xu, Ziheng Xi, Yuhui Fu, and Zongqing Lu. Being-h0.7: A latent world-action model from egocentric videos.arXiv preprint arXiv:2605.00078, 2026. 18

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion

    Jiangran Lyu, Kai Liu, Xuheng Zhang, Haoran Liao, Yusen Feng, Wenxuan Zhu, Tingrui Shen, Jiayi Chen, Jiazhao Zhang, Yifei Dong, Wenbo Cui, Senmao Qi, Shuo Wang, Yixin Zheng, Mi Yan, Xuesong Shi, Haoran Li, Dongbin Zhao, Ming-Yu Liu, Zhizheng Zhang, Li Yi, Yizhou Wang, and He Wang. LDA-1b: Scaling latent dynamics action model via universal embodied data in...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  24. [24]

    LARY: A Latent Action Representation Yielding Benchmark for Generalizable Vision-to-Action Alignment

    Dujun Nie, Fengjiao Chen, Qi Lv, Jun Kuang, Xiaoyu Li, Xuezhi Cao, and Xunliang Cai. Lary: A latent action representation yielding benchmark for generalizable vision-to-action alignment.arXiv preprint arXiv:2604.11689, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    Qwen3.6-27B: Flagship-level coding in a 27B dense model, April 2026

    Qwen Team. Qwen3.6-27B: Flagship-level coding in a 27B dense model, April 2026

    work page 2026

  26. [26]

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julie...

    work page 2025

  27. [27]

    World guidance: World modeling in condition space for action generation

    Yue Su, Sijin Chen, Haixin Shi, Mingyu Liu, Zhengshen Zhang, Ningyuan Huang, Weiheng Zhong, Zhengbang Zhu, Yuxiao Liu, and Xihui Liu. World guidance: World modeling in condition space for action generation. InICML, 2026

    work page 2026

  28. [28]

    Masked depth modeling for spatial perception

    Bin Tan, Changjian Sun, Xiage Qin, Hanat Adai, Zelin Fu, Tianxiang Zhou, Han Zhang, Yinghao Xu, Xing Zhu, Yujun Shen Shen, and Nan Xue. Masked depth modeling for spatial perception. https://technology.robbyant.com/lingbot-depth, 2026

    work page 2026

  29. [29]

    Towards human-like manipulation through RL-augmented teleoperation and mixture-of-dexterous-experts VLA.arXiv preprint arXiv:2603.08122, 2026

    Tutian Tang, Xingyu Ji, Wanli Xing, Ce Hao, Wenqiang Xu, Lin Shao, Cewu Lu, Qiaojun Yu, Jiangmiao Pang, and Kaifeng Zhang. Towards human-like manipulation through RL-augmented teleoperation and mixture-of-dexterous-experts VLA.arXiv preprint arXiv:2603.08122, 2026

    work page arXiv 2026

  30. [30]

    Gemini Robotics: Bringing AI into the Physical World

    Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, et al. Gemini Robotics: Bringing AI into the physical world.arXiv preprint arXiv:2503.20020, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Gen-1: Scaling embodied foundation models to mastery.Generalist AI Blog, 2026

    Generalist Team. Gen-1: Scaling embodied foundation models to mastery.Generalist AI Blog, 2026. https://generalistai.com/blog/gen-1

    work page 2026

  32. [32]

    GR00T N1.6: An improved open foundation model for generalist humanoid robots

    NVIDIA GEAR Team. GR00T N1.6: An improved open foundation model for generalist humanoid robots. https://research. nvidia.com/labs/gear/gr00t-n1_6/, 2025

    work page 2025

  33. [33]

    Qwen-robotmanip technical report: Alignment unlocks scale for robotic manipulation foundation models

    Qwen Team. Qwen-robotmanip technical report: Alignment unlocks scale for robotic manipulation foundation models. 2026

    work page 2026

  34. [34]

    Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments

    Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, Sicheng Xie, Yitao Liu, Junhao Chen, Zhixuan Liang, Jie Zhang, Xintong Hu, Xuhong Huang, Pei Lin, Junyang Lin, Dayiheng Liu, Shuai Bai, Jingren Zhou, Jiazhao Zhang, Haoqi Yuan, Gengze Zhou, Hang Yin, Ye Wang, Yiyang Huang, Zixing Lei, Wujian Peng, Delin Chen, et al. Qwen-VLA: Unifying vision-language-action...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  35. [35]

    Vision-centric activation and coordination for multimodal large language models.arXiv preprint arXiv:2510.14349, 2025

    Yunnan Wang, Fan Lu, Kecheng Zheng, Ziyuan Huang, Ziqiang Li, Wenjun Zeng, and Xin Jin. Vision-centric activation and coordination for multimodal large language models.arXiv preprint arXiv:2510.14349, 2025

    work page arXiv 2025

  36. [36]

    The Great March 100: 100 detail-oriented tasks for evaluating embodied ai agents, 2026

    Ziyu Wang, Chenyuan Liu, Yushun Xiang, Runhao Zhang, Qingbo Hao, Hongliang Lu, Houyu Chen, Zhizhong Feng, Kaiyue Zheng, Dehao Ye, Xianchao Zeng, Xinyu Zhou, Boran Wen, Jiaxin Li, Mingyu Zhang, Kecheng Zheng, Qian Zhu, Ran Cheng, and Yong-Lu Li. The Great March 100: 100 detail-oriented tasks for evaluating embodied ai agents, 2026

    work page 2026

  37. [37]

    Videorope: What makes for good video rotary position embedding? InInt

    Xilin Wei, Xiaoran Liu, Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Jian Tong, Haodong Duan, Qipeng Guo, Jiaqi Wang, et al. Videorope: What makes for good video rotary position embedding? InInt. Conf. Mach. Learn., 2025

    work page 2025

  38. [38]

    A Pragmatic VLA Foundation Model

    Wei Wu, Fan Lu, Yunnan Wang, Shuai Yang, Shi Liu, Fangjing Wang, Shuailei Ma, He Sun, Yong Wang, Zhenqi Qiu, Houlong Xiong, Ziyu Wang, Shuai Zhou, Yiyu Ren, Kejia Zhang, Hui Yu, Jingmei Zhao, Qian Zhu, Ran Cheng, Yong-Lu Li, Yongtao Huang, Xing Zhu, Yujun Shen, and Kecheng Zheng. A pragmatic vla foundation model.arXiv preprint arXiv:2601.18692v1, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  39. [39]

    Hy-embodied-0.5-x: An enhanced embodied foundation model for real-world agents

    Tencent Robotics X and HY Vision Team. Hy-embodied-0.5-x: An enhanced embodied foundation model for real-world agents. 2026

    work page 2026

  40. [40]

    Omnistream: Mastering perception, reconstruction and action in continuous streams.arXiv preprint arXiv:2603.12265, 2026

    Yibin Yan, Jilan Xu, Shangzhe Di, Haoning Wu, and Weidi Xie. Omnistream: Mastering perception, reconstruction and action in continuous streams.arXiv preprint arXiv:2603.12265, 2026

    work page arXiv 2026

  41. [41]

    Magma: A foundation model for multimodal AI agents

    Jianwei Yang, Reuben Tan, Qianhui Wu, Ruijie Zheng, Baolin Peng, Yongyuan Liang, Yu Gu, Mu Cai, Seonghyeon Ye, Joel Jang, et al. Magma: A foundation model for multimodal AI agents. InIEEE Conf. Comput. Vis. Pattern Recog., pages 14203–14214, 2025. 19

    work page 2025

  42. [42]

    ABot-M0: VLA Foundation Model for Robotic Manipulation with Action Manifold Learning

    Yandan Yang, Shuang Zeng, Tong Lin, Xinyuan Chang, Dekang Qi, Junjin Xiao, Haoyun Liu, Ronghan Chen, Yuzhi Chen, Dongjie Huo, et al. Abot-m0: Vla foundation model for robotic manipulation with action manifold learning.arXiv preprint arXiv:2602.11236, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  43. [43]

    StarVLA-$\alpha$: Reducing Complexity in Vision-Language-Action Systems

    Jinhui Ye, Ning Gao, Senqiao Yang, Jinliang Zheng, Zixuan Wang, Yuxin Chen, Pengguang Chen, Yilun Chen, Shu Liu, and Jiaya Jia. Starvla- alpha: Reducing complexity in vision-language-action systems.arXiv preprint arXiv:2604.11757, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    SAMoE-VLA: A scene adaptive mixture-of-experts vision-language-action model for autonomous driving.arXiv preprint arXiv:2603.08113, 2026

    Zihan You, Hongwei Liu, Chenxu Dang, Zhe Wang, Sining Ang, Aoqi Wang, and Yan Wang. SAMoE-VLA: A scene adaptive mixture-of-experts vision-language-action model for autonomous driving.arXiv preprint arXiv:2603.08113, 2026

    work page arXiv 2026

  45. [45]

    ForceVLA: Enhancing VLA models with a force-aware MoE for contact-rich manipulation

    Jiawen Yu, Hairuo Liu, Qiaojun Yu, Jieji Ren, Ce Hao, Haitong Ding, Guangyu Huang, Guofan Huang, Yan Song, Panpan Cai, Cewu Lu, and Wenqiang Zhang. ForceVLA: Enhancing VLA models with a force-aware MoE for contact-rich manipulation. arXiv preprint arXiv:2505.22159, 2025

    work page arXiv 2025

  46. [46]

    Wall-OSS-0.5 Technical Report

    Ryan Yu, Pushi Zhang, Starrick Liu, Brae Liu, Miracle Kang, Shalfun Li, Lights Shi, Ellie Ma, Ping Yang, Chris Pan, Jerry Chen, Dongxiu Liu, Rain Sun, Miles Guo, Byron Zhang, Hugo Zhou, Zach Xu, Vincent Chen, Harrison Huang, James Wang, Dance Kuzi, Andy Zhai, Hang Su, Roy Gan, Lucy Liang, Hao Wang, and Qian Wang. Wall-oss-0.5 technical report.arXiv prepri...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  47. [47]

    Igniting VLMs toward the embodied space.arXiv preprint arXiv:2509.11766, 2025

    Andy Zhai, Brae Liu, Bruno Fang, Chalse Cai, Ellie Ma, Ethan Yin, Hao Wang, Hugo Zhou, James Wang, Lights Shi, et al. Igniting VLMs toward the embodied space.arXiv preprint arXiv:2509.11766, 2025

    work page arXiv 2025

  48. [48]

    AtomicVLA: Unlocking the potential of atomic skill learning in robots.arXiv preprint arXiv:2603.07648, 2026

    Likui Zhang, Tao Tang, Zhihao Zhan, Xiuwei Chen, Zisheng Chen, Jianhua Han, Jiangtong Zhu, Pei Xu, Hang Xu, Hefeng Wu, Liang Lin, and Xiaodan Liang. AtomicVLA: Unlocking the potential of atomic skill learning in robots.arXiv preprint arXiv:2603.07648, 2026

    work page arXiv 2026

  49. [49]

    GEM: Generative Supervision Helps Embodied Intelligence

    Ruowen Zhao, Bangguo Li, Zuyan Liu, Yinan Liang, Junliang Ye, Fangfu Liu, Diankun Wu, Zhengyi Wang, Xumin Yu, Yongming Rao, Han Hu, and Jun Zhu. GEM: Generative supervision helps embodied intelligence.arXiv preprint arXiv:2605.28548, 2026. 20

    work page internal anchor Pith review Pith/arXiv arXiv 2026