Pith. sign in

REVIEW 2 major objections 7 minor 50 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

Robot vision model infers camera pose itself, no calibration needed

2026-07-07 12:19 UTC pith:KJXS3467

load-bearing objection CamVLA cleanly removes the calibration requirement for view-robust VLA by predicting camera-centric actions and a learned hand-eye rotation from monocular RGB. The core math is correct and the gains are real, but the evaluation range is narrow and the noise-tolerance ablation has a gap that matters. the 2 major comments →

arxiv 2607.05396 v1 pith:KJXS3467 submitted 2026-07-06 cs.CV cs.AIcs.LGcs.RO

From Fixed to Free Cameras: Calibration-Free View-Robust Vision-Language-Action Model

Wenhao Li , Xueying Jiang , Quanhao Qian , Deli Zhao , Shijian Lu , Gongjie Zhang , Ran Xu This is my paper
classification cs.CV cs.AIcs.LGcs.RO
keywords vision-language-action modelviewpoint robustnesshand-eye calibrationcamera-centric actionrobot manipulationcalibration-free
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.

The paper argues that current vision-language-action (VLA) models fail when a camera is moved because they implicitly try to learn a fixed mapping from camera image to robot-base coordinates. CamVLA replaces that implicit mapping with an explicit two-part prediction: an action head that outputs movement in the camera's own coordinate frame, and a geometric head that regresses the hand-eye matrix (the rigid transform relating the camera to the robot base) directly from a single RGB image. A deterministic geometric transformation then composes these two outputs into the robot's base-frame action. The central claim is that this factorization enables the policy to self-localize and maintain robust manipulation under unseen camera viewpoints without any external calibration, depth, or multi-view input at deployment. The paper demonstrates that this approach improves success rates across diverse unseen viewpoints in both simulation and real-world robot experiments, with negligible computational overhead.

Core claim

The core discovery is that decoupling a VLA policy into camera-centric action prediction and learned hand-eye matrix regression from monocular RGB eliminates the need for externally provided camera extrinsics at deployment. By parameterizing actions natively in the camera frame, the visual-to-action mapping becomes pose-independent; by making the hand-eye matrix an explicit network output, viewpoint variability is absorbed into a single learned 6-DoF pose rather than entangled in policy weights. The paper further shows that delta-action execution is mathematically independent of the hand-eye translation component, confining all viewpoint-sensitive execution error to the rotation component, a

What carries the argument

hand-eye matrix

Load-bearing premise

The approach assumes that a single monocular RGB image contains enough information for the geometric head to accurately regress the camera-to-robot rotation matrix at deployment. At a 15-degree viewpoint offset, rotation error reaches 9.39 degrees, which is close to the roughly 12-degree tolerance boundary the policy can withstand before performance degrades sharply.

What would settle it

If the geometric head's rotation regression degrades faster than the closed-loop policy's tolerance under more extreme or out-of-distribution viewpoints than the tested range (up to 15 degrees), the entire framework's advantage over baselines would erode.

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

If this is right

  • Robot deployment in unstructured environments becomes simpler: no hand-eye calibration step is needed when cameras are bumped, remounted, or hand-held.
  • The factorization principle—separating pose-invariant action generation from learned geometric grounding—could extend to other sensor perturbations beyond camera viewpoint, such as base-frame drift on mobile platforms.
  • If the geometric head's rotation regression could be improved (e.g., via temporal filtering or online adaptation), the framework's tolerance envelope for viewpoint shifts would widen proportionally.

Where Pith is reading between the lines

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

  • The mathematical independence of delta-action execution from hand-eye translation (Eqs. 7-10) means the translation component of the geometric head is purely a training-time regularizer with zero deployment impact—a structural property that could be exploited to simplify or regularize training further.
  • The framework's reliance on a single monocular RGB image for self-localization sets a ceiling on achievable viewpoint range; combining the learned geometric head with even crude inertial or depth signals could extend robustness beyond the tested 15-degree offset without sacrificing the calibration-free property.
  • The analogy to biological egocentric/allocentric visual processing suggests the factorization may be a general principle for sensorimotor learning, not specific to camera-robot hand-eye transforms.

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

2 major / 7 minor

Summary. The paper introduces CamVLA, a VLA framework that decouples manipulation policy into a camera-centric action head (predicting end-effector deltas in the camera frame) and a geometric head (regressing the 6-DoF hand-eye matrix from monocular RGB), composed via a deterministic rigid-body transformation (Eqs. 3-4). The central claim is that this factorization enables calibration-free, depth-free, single-view viewpoint robustness without external camera information at deployment. The approach is evaluated in simulation (RLBench, 6 tasks, dense viewpoint grid) and real-world (5 tasks, 3 cameras, 3 offsets) with two VLA backbones (π0, GR00T N1.7). The mathematical derivation (Supp. Sec. B) showing translation invariance of delta actions is correct and clearly presented. The experimental evaluation demonstrates consistent improvements over baselines under unseen viewpoints.

Significance. The paper addresses a practically important problem: viewpoint robustness of VLA policies without requiring calibrated extrinsics at deployment. The factorization into camera-centric actions and learned geometric grounding is a clean and well-motivated design. Strengths include: (1) a correct and clearly presented mathematical derivation showing that delta-action execution is invariant to hand-eye translation (Eqs. 7-10, Supp. Sec. B), which is a non-trivial property that simplifies the regression target; (2) comprehensive experiments across two strong VLA backbones, six simulation tasks, and five real-world tasks with ablations on training density, state/action representations, feature sources, and pose representations; (3) minimal computational overhead (Table 4: +6.3M params, +1ms latency); (4) an honest limitations section acknowledging struggles under extreme viewpoint changes. The approach is falsifiable: the geometric head's rotation accuracy is measured directly (Tables 3, 5), and the noise-tolerance ablation (Supp. Fig. 5) provides a concrete failure threshold.

major comments (2)
  1. Supp. Fig. 5 and the associated noise-tolerance ablation: The rotation-noise robustness experiment feeds the action head canonical-view images while injecting noise only into the ground-truth extrinsics. At deployment on an unseen viewpoint, both the action head and geometric head process the same out-of-distribution image, so their errors are correlated. The paper itself provides evidence that this coupled degradation is already occurring: at 15° real-world offset, CamVLA achieves 29.3% success (π0 backbone, Table 2) with 9.39° rotation error (Table 3), yet Supp. Fig. 5 suggests that with GT extrinsics plus ~9° injected noise, success should be approximately 45-50% (interpolating between 58.7% at 5° and 36.0% at 12°). The gap between the ablation's prediction and the actual deployment result suggests the action head's own degradation under viewpoint shift compounds with rotation error.
  2. Table 2 and Table 3, 15° offset results: At 15° real-world offset, the translation error of the geometric head reaches 27.16 cm (Table 3). While the paper correctly argues that delta-action execution is translation-invariant (Eqs. 7-10), the rotation error of 9.39° is close to the tolerance boundary suggested by Supp. Fig. 5 (where performance drops sharply between 12° and 15° of injected noise). The paper acknowledges this in the limitations section but does not quantify the margin or discuss what happens beyond 15°. Given that the method is motivated by deployment scenarios with 'hand-held, drifting, or remounted cameras' (§1), the tested range of ±15° is relatively narrow. The paper should either (a) test more extreme offsets to establish the failure boundary, or (b) explicitly scope the claim to 'typical camera perturbations' rather than general viewpoint robustness. The current phr
minor comments (7)
  1. §3.2: The phrase 'this consistent spatial relationship prevents visual-action confusion' is stated without rigorous justification. Consider grounding this claim more concretely, e.g., by referencing the ablation in Table 7 that shows the camera-frame action output alone improves success from 33.2% to 51.9%.
  2. Table 5: The notation 'CamVLA†' is used for the ground-truth extrinsics variant, but this notation is not defined until Table 7. Define it at first use.
  3. Table 9: The VLM Backbone configuration achieves 53.5% success (higher than the default Image Encoder's 51.4%) but is rejected due to geometric instability at unseen viewpoints. This is a non-trivial design decision; a brief mention of the specific failure mode (e.g., 'geometric discontinuities and large localization spikes') with a quantitative example would strengthen the justification.
  4. §4.2: The real-world evaluation uses 5 training viewpoints and tests on 3 of those cameras with angular offsets. It would help to clarify whether the 5 training cameras span a range of azimuthal angles or are clustered, as this affects the interpretation of generalization results.
  5. Supp. Sec. A: The geometric loss weight λ=0.1 is stated without justification. A brief note on how this was selected (e.g., grid search, sensitivity analysis) would be useful.
  6. Figure 2: The architecture diagram could be clearer about where the geometric head's features come from (image encoder vs. VLM backbone). This is addressed in Supp. Table 9 but not visible in the main figure.
  7. The paper uses 'zero-shot generalization' to describe evaluation on unseen viewpoints within the interpolation range of training viewpoints (e.g., testing at 5° intervals between 15° training intervals). Consider clarifying that this is interpolation within the training distribution rather than extrapolation, to avoid confusion with the standard zero-shot transfer setting.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and the insightful quantitative analysis of the gap between our noise-tolerance ablation and real-world deployment results. Both major comments are well-taken and will lead to revisions in the manuscript.

read point-by-point responses

  1. Referee: Supp. Fig. 5 noise-tolerance ablation: The rotation-noise robustness experiment feeds the action head canonical-view images while injecting noise only into the ground-truth extrinsics. At deployment, both heads process the same OOD image, so errors are correlated. The gap between the ablation's prediction (~45-50% at 9° noise) and actual 15° deployment result (29.3%) suggests compounding degradation.

    Authors: The referee is correct on both the structural limitation of the ablation and the quantitative discrepancy it produces. We acknowledge that Supp. Fig. 5 isolates rotation noise from action-head visual degradation, and therefore overestimates deployment performance under coupled OOD conditions. The referee's back-of-envelope calculation—predicting ~45-50% success from the ablation versus the observed 29.3% at 15° real-world offset—is a compelling demonstration that the action head's own viewpoint-induced degradation compounds with geometric-head rotation error. We do not have an alternative explanation for this gap; the compounding-error interpretation is the most natural one. We will revise the manuscript to: (1) explicitly state that the Supp. Fig. 5 ablation uses canonical-view images and therefore does not capture coupled degradation; (2) add the referee's quantitative comparison as a concrete illustration of this compounding effect; and (3) note that our simulation experiments (Table 1, viewpoints up to ±90°) do reflect coupled degradation since both heads process unseen views, but the specific noise-tolerance ablation does not. We agree this is an important caveat on the interpretive scope of Supp. Fig. 5. revision: yes

  2. Referee: Table 2 and Table 3, 15° offset results: The tested range of ±15° is relatively narrow for the stated motivation of 'hand-held, drifting, or remounted cameras.' The paper should either (a) test more extreme offsets to establish the failure boundary, or (b) explicitly scope the claim to 'typical camera perturbations' rather than general viewpoint robustness.

    Authors: We agree that the real-world evaluation range of ±15° is narrower than what the introduction's framing might suggest. We chose this range because, as noted in §4.2, beyond 15° both the baseline and our model perform poorly on real hardware, making further evaluation impractical for demonstrating useful manipulation. However, we concede that this practical justification does not fully address the gap between the motivation (hand-held, drifting, remounted cameras) and the tested range. We will adopt option (b): we will revise the manuscript to scope the real-world claims to 'typical camera perturbations' (e.g., sensor drift, minor repositioning) rather than general viewpoint robustness, and will make clear that the broader viewpoint range (±90°) is validated only in simulation. We will also add a brief discussion of what the simulation results at extreme angles (Table 1, averaged over viewpoints up to ±90°) do and do not tell us about real-world failure boundaries, noting that sim-to-real gaps in visual appearance may further constrain the effective range. We note that our simulation experiments do cover a much wider range of viewpoints (training at 15° intervals from −90° to +90°, testing on all intermediate angles), so the method is evaluated under extreme viewpoint shifts in at least one domain. revision: yes

Circularity Check

0 steps flagged

No circularity found: the geometric derivation is standard rigid-body kinematics, and the central claim is validated against external benchmarks.

full rationale

The paper's core derivation (Eqs. 3-4, 7-10) is a standard rigid-body kinematics result: delta actions (free vectors) transform linearly under rotation, and the hand-eye translation cancels exactly for relative displacements. This is textbook material (cited to Murray, Li & Sastry [49]; Siciliano et al. [48]; Tsai & Lenz [44]; Shiu & Ahmad [45]) and is not a self-cited or self-defined result. The hand-eye matrix formulation draws from classical robotics literature with no author overlap. The VLA baselines (π0 [5], GR00T N1.7 [4]) are external models. The loss function L = L_act + λL_ext with λ=0.1 is a standard hyperparameter choice, not a fitted input renamed as a prediction. The paper's central claim—that the factorization enables calibration-free viewpoint robustness—is tested against external benchmarks (RLBench simulation, real-world Franka robot) with success rates and hand-eye error metrics that are independently measurable and not forced by construction. The ablation in Table 7 (comparing self-predicted vs. ground-truth extrinsics) shows a 0.9% gap, which is an empirical finding rather than a tautological consequence of the formulation. The rotation-noise tolerance ablation (Supp. Fig. 5) uses ground-truth extrinsics with injected noise as a controlled experiment, not as a prediction claimed to match deployment conditions. No step in the derivation chain reduces to its own inputs by definition, fit, or self-citation. The paper is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

CamVLA introduces no new physical entities, particles, forces, or dimensions. The 'Geometric Head' and 'Action Head' are architectural components, not postulated physical objects. The hand-eye matrix is a standard robotics concept. The free parameters are standard hyperparameters (loss weight, network size, training configuration) rather than constants fitted to make a derivation work. The axioms are domain assumptions about the sufficiency of monocular RGB for pose regression and the tolerance of closed-loop policies to rotation noise — both empirically tested but not theoretically guaranteed.

free parameters (4)
  • λ (geometric loss weight) = 0.1
    Weighting factor for the hand-eye regression loss L_ext relative to the action loss L_act. Set to 0.1 without systematic justification for this specific value.
  • Geometric Head architecture (hidden dim, layers) = 1024, 3 layers
    The MLP architecture for the geometric head is specified as 3-layer with 1024 hidden dim and GELU activations. These are design choices without ablation over alternative architectures.
  • Training viewpoint interval (simulation) = 15°
    The default training viewpoint sampling interval of 15° is chosen for the main results. Table 5 ablates 30° and 45° but 15° is the primary configuration.
  • Number of training viewpoints (real-world) = 5
    Five camera perspectives are used for real-world training data collection. This number is chosen as providing 'sufficient viewpoint diversity' without systematic justification.
axioms (4)
  • domain assumption A single monocular RGB image contains sufficient geometric information to regress the hand-eye rotation matrix R_t with accuracy adequate for closed-loop control.
    This is the foundational assumption of the Geometric Head. The paper provides empirical evidence (Table 3: 1.41° rotation error at 15° training interval in simulation) but no theoretical guarantee. It enters in Sec. 3.3 where the geometric head is introduced.
  • domain assumption Closed-loop VLA policies have inherent tolerance to small rotation errors in the hand-eye matrix, such that errors under ~12° do not catastrophically degrade manipulation.
    This is empirically supported by the noise injection experiment (Supp. Fig. 5) but is an assumption about the robustness of the underlying policy rather than a proven property. It enters in Sec. 4.2 and Supp. Sec. B/C.
  • domain assumption Delta (relative) actions are a sufficient action representation for the manipulation tasks evaluated.
    The translation invariance result (Eqs. 7-10) holds only for delta actions. If absolute actions were needed, the large translation errors (Table 3: 27.16 cm at 15°) would be catastrophic. The paper acknowledges this in Sec. 3.4: 'we still regress τ_t to enhance geometric grounding... and to support potential absolute-action variants.'
  • standard math The rigid-body hand-eye transform T_t ∈ SE(3) adequately models the camera-to-robot-base relationship during deployment.
    Standard assumption from robotics kinematics (Tsai & Lenz [44]). Invoked throughout Sec. 3.

pith-pipeline@v1.1.0-glm · 21442 in / 3774 out tokens · 176690 ms · 2026-07-07T12:19:56.102605+00:00 · methodology

0 comments
read the original abstract

Real-world robot deployment rarely maintains the training-stage camera setup, where cameras often experience repositioning or remounting depending on actual scenarios. Existing view-robust Vision-Language-Action (VLA) policies tolerate such camera variations only when the camera extrinsics are explicitly provided, making them fragile and hard to use especially when view robustness is critical. We argue that the policy should not be told where the camera is, but rather figure it out by itself. To this end, we introduce Camera-Centric VLA (CamVLA), a new VLA model that decouples manipulation controls from camera geometry by predicting (i) a camera-centric end-effector action expressed in the local camera frame, and (ii) a 6-DoF hand-eye matrix relating cameras to the robot base. A deterministic geometric transformation composes the two predictions into a robot base-frame action. This disentangles how I should move in pose-independent camera-centric action generation from where I am looking from in camera-perspective geometric grounding. The resulting policy is calibration-free, depth-free, and single-view, requiring only a single monocular RGB image as the visual observation and task instruction at deployment. Evaluations in both simulation and real-world robot data show that CamVLA consistently improves success rates across diverse unseen viewpoints. Project page: https://alibaba-damo-academy.github.io/CamVLA/.

Figures

Figures reproduced from arXiv: 2607.05396 by Deli Zhao, Gongjie Zhang, Quanhao Qian, Ran Xu, Shijian Lu, Wenhao Li, Xueying Jiang.

Figure 1
Figure 1. Figure 1: The Viewpoint Trap in VLAs. Conven￾tional VLA (e.g., π0) trained on a single viewpoint exhibits extreme spatial brittleness, where a mere 15◦ camera shift drops success rates to 6.3%. In this work, we argue that the policy should not be told where the camera is, but rather should figure it out by itself. This self-localizing capa￾bility finds a natural analogue in human cogni￾tion: human vision-guided mani… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the CamVLA Architecture. Our CamVLA predicts the local camera-centric action and the 6-DoF hand-eye pose in parallel, which are then combined via a deterministic geometric transformation to execute the base-frame action. 2 Related Work Vision-Language-Action Models. The pursuit of generalist robot policies has catalyzed the develop￾ment of VLAs [14, 15, 16, 17, 18], which adapt VLMs [19, 20, 21… view at source ↗
Figure 3
Figure 3. Figure 3: Simulation camera configuration. The training set (red cameras) covers discrete view￾points, while evaluation (green cameras) is con￾ducted on a dense set of unseen viewpoints. Experimental Setup. We evaluate CamVLA on the RLBench benchmark [6]. To evaluate view￾point robustness, we utilize the front camera and rotate it around the robot base from −90◦ to 90◦ at 5 ◦ intervals to generate a diverse set of v… view at source ↗
Figure 4
Figure 4. Figure 4: Real-world experimental setup. Multi￾view camera configuration used to verify viewpoint robustness in physical environments. Experimental Setup. Our real-world setup uses a Franka Research 3 robot arm with a parallel gripper. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Robustness under artificial extrinsic rotation noise. Success rate (%) under varying levels of random rotation noise (ranging from 0 ◦ to 45◦ ) applied to the ground-truth hand-eye rotation matrix during execution. comparable to the default 6-DoF configuration (51.4% success rate and 1.4◦ ). This demonstrates that additionally predicting the translation component does not degrade task success or rotation e… view at source ↗
Figure 6
Figure 6. Figure 6: Simulation tasks on RLBench benchmark. We evaluate CamVLA across a diverse set of manipulation tasks, requiring both high-level semantic understanding and precise low-level control. -90° -75° -60° -45° -30° -15° 0° 15° 30° 45° 60° 75° 90° Training View Testing View -85° -80° -70° -65° -55° -50° -40° -35° -25° -20° -10° -5° 85° 80° 70° 65° 55° 50° 40° 35° 25° 20° 10° 5° [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of training and testing viewpoints on RLBench. The training set consists of views sampled at 15◦ intervals (top row), while the testing set covers a dense range of unseen viewpoints (middle and bottom rows) to evaluate the zero-shot generalization capability of CamVLA. -90° 90° -75° -60° -45° -30° -15° 0° 15° 30° 45° 60° 75° -10°-5° 85° 80° 70° 65° 55° 50° 40° 35° 25° 20° 10° 5° -85° -80° -70… view at source ↗
Figure 8
Figure 8. Figure 8: Detailed range of camera viewpoints in simulation. (a) Training viewpoints sampled at 30◦ intervals and (b) training viewpoints sampled at 45◦ intervals. Red and green cameras represent training and unseen testing viewpoints, respectively. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Real-world evaluation tasks. Five representative manipulation tasks involving diverse objects and interaction requirements. Cam 1 Cam 2 Cam 3 Cam 4 Cam 5 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Real-world training camera viewpoints. Five third-person training placements (Cam 1–Cam 5) used for demonstration collection. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Real-world unseen testing viewpoints. Cam 2, Cam 3, and Cam 4 are horizontally rotated by 5 ◦ , 10◦ , and 15◦ from their canonical 0 ◦ training position to form unseen testing views. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison between π0 and CamVLA on RLBench under unseen cameras. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison between π0 and CamVLA on real-world robot experiments under repositioned cameras [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Visualization of real-world experiments under dynamically hand-held moving cameras. We visualize the robot’s execution and dynamic hand-eye pose tracking when the third￾person camera is continuously moved by a human operator during deployment. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Failure cases. We illustrate common failure modes such as boundary objects, actions exceeding the robot’s physical workspace, and self-occlusion. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] 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

50 extracted references · 50 canonical work pages · 7 internal anchors

  1. [1]

    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. In RSS, 2022

    work page 2022

  2. [2]

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

    work page 2025

  3. [3]

    Intelligence, B

    P. Intelligence, B. Ai, A. Amin, R. Aniceto, A. Balakrishna, G. Balke, K. Black, G. Bokinsky, S. Cao, T. Charbonnier, et al. π∗ 0.7: a steerable generalist robotic foundation model with emergent capabilities. 2026

    work page 2026

  4. [4]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. 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]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. π0: A vision-language-action flow model for general robot control. InRSS, 2024

    work page 2024

  6. [6]

    James, Z

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison. RLBench: The robot learning benchmark & learning environment.IEEE Robotics and Automation Letters, 5(2):3019–3026, 2020

    work page 2020

  7. [7]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. OpenVLA: An open-source vision-language-action model. InCoRL, 2024

    work page 2024

  8. [8]

    Zhang, H

    T. Zhang, H. Duan, H. Hao, Y . Qiao, J. Dai, and Z. Hou. Grounding actions in camera space: Observation-centric vision-language-action policy. InAAAI, volume 40, pages 18782–18790, 2026

    work page 2026

  9. [9]

    Jiang, J

    T. Jiang, J. Ji, X. Tan, J. Fang, A. Bhattad, V . Guizilini, and M. R. Walter. Do you know where your camera is? view-invariant policy learning with camera conditioning.arXiv preprint arXiv:2510.02268, 2025

    work page arXiv 2025

  10. [10]

    Zhang, Y

    J. Zhang, Y . Chen, Y . Xu, Z. Huang, Y . Zhou, Y .-J. Yuan, X. Cai, G. Huang, X. Quan, H. Xu, et al. 4D-VLA: Spatiotemporal vision-language-action pretraining with cross-scene calibration. arXiv preprint arXiv:2506.22242, 2025

    work page arXiv 2025

  11. [11]

    H. Heo, S. Woo, S. M. Kim, J. Kim, J. Lee, Y . Lee, and Y . M. Kim. AnyCamVLA: Zero- shot camera adaptation for viewpoint robust vision-language-action models.arXiv preprint arXiv:2603.05868, 2026

    work page arXiv 2026

  12. [12]

    M. A. Goodale and A. D. Milner. Separate visual pathways for perception and action.Trends in neurosciences, 15(1):20–25, 1992

    work page 1992

  13. [13]

    J. D. Crawford, W. P. Medendorp, and J. J. Marotta. Spatial transformations for eye–hand coordination.Journal of neurophysiology, 2004

    work page 2004

  14. [14]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. InCoRL, pages 2165–2183, 2023

    work page 2023

  15. [15]

    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. InICML, 2023

    work page 2023

  16. [16]

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakr- ishnan, K. Hausman, et al. Do as I can, not as I say: Grounding language in robotic affordances. InCoRL, 2022. 9

    work page 2022

  17. [17]

    Open X-Embodiment: Robotic Learning Datasets and RT-X Models

    O.-E. Collaboration, A. O’Neill, A. Rehman, A. Gupta, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, et al. Open X-Embodiment: Robotic learning datasets and RT-X models.arXiv preprint arXiv:2310.08864, 1(2), 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    $\pi^{*}_{0.6}$: a VLA That Learns From Experience

    P. Intelligence, A. Amin, R. Aniceto, A. Balakrishna, K. Black, K. Conley, G. Connors, J. Darpinian, K. Dhabalia, J. DiCarlo, et al. π∗ 0.6: a VLA that learns from experience.arXiv preprint arXiv:2511.14759, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning.NeurIPS, 36:34892–34916, 2023

    work page 2023

  20. [20]

    J. Li, D. Li, S. Savarese, and S. Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InICML, pages 19730–19742, 2023

    work page 2023

  21. [21]

    D. Zhu, X. Shen, X. Li, M. Elhoseiny, et al. MiniGPT-4: Enhancing vision-language under- standing with advanced large language models. InICLR, volume 2024, pages 18378–18394, 2024

    work page 2024

  22. [22]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y . Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y . Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin. Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    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. InRSS, 2024

    work page 2024

  24. [24]

    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.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

    work page 2025

  25. [25]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al. π0.5: a vision-language-action model with open-world generalization. InCoRL, 2025

    work page 2025

  26. [26]

    H. Chen, J. Liu, C. Gu, Z. Liu, R. Zhang, X. Li, X. He, Y . Guo, C.-W. Fu, S. Zhang, et al. Fast-in-slow: A dual-system foundation model unifying fast manipulation within slow reasoning. arXiv preprint arXiv:2506.01953, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    W. Li, Q. Zhang, R. Zhai, L. Lin, and G. Wang. VLA models are more generalizable than you think: Revisiting physical and spatial modeling.arXiv preprint arXiv:2512.02902, 2025

    work page arXiv 2025

  28. [28]

    Shridhar, L

    M. Shridhar, L. Manuelli, and D. Fox. Perceiver-Actor: A multi-task transformer for robotic manipulation. InCoRL, pages 785–799, 2023

    work page 2023

  29. [29]

    H. Zhen, X. Qiu, P. Chen, J. Yang, X. Yan, Y . Du, Y . Hong, and C. Gan. 3D-VLA: A 3D vision-language-action generative world model. InICML, 2024

    work page 2024

  30. [30]

    Y . Jia, J. Liu, S. Chen, C. Gu, Z. Wang, L. Luo, L. Lee, P. Wang, Z. Wang, R. Zhang, et al. Lift3D foundation policy: Lifting 2D large-scale pretrained models for robust 3D robotic manipulation.arXiv preprint arXiv:2411.18623, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    Zhang, H

    Z. Zhang, H. Li, Y . Dai, Z. Zhu, L. Zhou, C. Liu, D. Wang, F. E. Tay, S. Chen, Z. Liu, et al. From spatial to actions: Grounding vision-language-action model in spatial foundation priors. arXiv preprint arXiv:2510.17439, 2025

    work page arXiv 2025

  32. [32]

    T.-W. Ke, N. Gkanatsios, and K. Fragkiadaki. 3D Diffuser Actor: Policy diffusion with 3D scene representations. InCoRL, 2024

    work page 2024

  33. [33]

    S. Deng, M. Yan, Y . Zheng, J. Su, W. Zhang, X. Zhao, H. Cui, Z. Zhang, and H. Wang. StereoVLA: Enhancing vision-language-action models with stereo vision.arXiv preprint arXiv:2512.21970, 2025. 10

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    S. Tian, B. Wulfe, K. Sargent, K. Liu, S. Zakharov, V . Guizilini, and J. Wu. View-invariant policy learning via zero-shot novel view synthesis. InCoRL, 2024

    work page 2024

  35. [35]

    Coholich, J

    J. Coholich, J. Wit, R. Azarcon, and Z. Kira. Sim2real image translation enables viewpoint- robust policies from fixed-camera datasets.arXiv preprint arXiv:2601.09605, 2026

    work page arXiv 2026

  36. [36]

    Y . Xu, J. Yang, X. Wang, Y . Chen, Z. Zhu, B. Fang, G. Huang, X. Chen, Y . Ye, Q. Zhang, et al. EgoDemoGen: Novel egocentric demonstration generation enables viewpoint-robust manipulation.arXiv preprint arXiv:2509.22578, 2025

    work page arXiv 2025

  37. [37]

    Vasudevan, S

    S. Vasudevan, S. Sagar, and R. Senanayake. Viewpoint-agnostic manipulation policies with strategic vantage selection.arXiv preprint arXiv:2506.12261, 2025

    work page arXiv 2025

  38. [38]

    J.-C. Pang, N. Tang, K. Li, Y . Tang, X.-Q. Cai, Z.-Y . Zhang, G. Niu, M. Sugiyama, and Y . Yu. Learning view-invariant world models for visual robotic manipulation. InICLR, 2025

    work page 2025

  39. [39]

    Jeong, J

    Y . Jeong, J. Chun, and T. Kim. Learning to act robustly with view-invariant latent actions.arXiv preprint arXiv:2601.02994, 2026

    work page arXiv 2026

  40. [40]

    M. Luo, Z. Xue, A. Dimakis, and K. Grauman. Viewpoint rosetta stone: Unlocking unpaired ego-exo videos for view-invariant representation learning. InCVPR, pages 15802–15812, 2025

    work page 2025

  41. [41]

    Z. Li, P. Qu, Y . Jia, S. Zhou, H. Ge, J. Cao, J. Zhou, G. Zhou, and J. Ma. Manivid-3D: Generalizable view-invariant reinforcement learning for robotic manipulation via disentangled 3D representations.IEEE Robotics and Automation Letters, 2026

    work page 2026

  42. [42]

    J. Q. Sun, H. Weng, X. Xing, C. M. Yeum, and M. Crowley. View invariant learning for vision-language navigation in continuous environments.IEEE Robotics and Automation Letters, 2026

    work page 2026

  43. [43]

    Abouzeid, M

    A. Abouzeid, M. Mansour, Q. Sun, Z. Sun, and D. Song. GeoAware-VLA: Implicit geometry aware vision-language-action model.arXiv preprint arXiv:2509.14117, 2025

    work page arXiv 2025

  44. [44]

    R. Y . Tsai, R. K. Lenz, et al. A new technique for fully autonomous and efficient 3 d robotics hand/eye calibration.IEEE Transactions on robotics and automation, 5(3):345–358, 1989

    work page 1989

  45. [45]

    Y . C. Shiu and S. Ahmad. Calibration of wrist-mounted robotic sensors by solving homogeneous transform equations of the form ax= xb. 1987

    work page 1987

  46. [46]

    Q. Qian, G. Zhao, G. Zhang, J. Wang, R. Xu, J. Gao, and D. Zhao. GP3: A 3D geometry-aware policy with multi-view images for robotic manipulation. InICRA, 2025

    work page 2025

  47. [47]

    B. Cai, Q. Liang, J. Li, S. Weng, Z. Zhang, T. Lin, X. Chen, W. Zhang, J. Mao, W. Xu, et al. Beyond viewpoint generalization: What multi-view demonstrations offer and how to synthesize them for robot manipulation?arXiv preprint arXiv:2603.26757, 2026

    work page arXiv 2026

  48. [48]

    Siciliano, L

    B. Siciliano, L. Sciavicco, L. Villani, and G. Oriolo.Robotics: modelling, planning and control. 2009

    work page 2009

  49. [49]

    R. M. Murray, Z. Li, and S. S. Sastry.A mathematical introduction to robotic manipulation. 2017

    work page 2017

  50. [50]

    K0jHdpnlKNMordptp9Im90XWy1I=

    I. A. Sucan, M. Moll, and L. E. Kavraki. The open motion planning library.IEEE Robotics & Automation Magazine, 19(4):72–82, 2012. 11 Supplementary Material This supplementary material covers the following details: • Additional Implementation details (Sec. A). • Additional Mathematical derivation (Sec. B). • Additional ablation studies (Sec. C). • Addition...

    work page 2012