Robotics · arXiv:2609.13812 v2 · submitted 12 Sep 2026
GeomVLA: Unifying Scene, Motion, and Action in 3D
The paper says
“We present GeomVLA, a Vision-Language-Action (VLA) model that unifies perception, latent scene motion prediction, and action generation within a shared robot-centric 3D coordinate frame.”
more of the abstract
Our approach lifts pretrained VLM features into spatially grounded 3D scene tokens using depth and camera calibration, while retaining the semantic representations learned during VLM pretraining. We further introduce a 3D Scene Trajectory Denoiser, a task-conditioned module that learns a latent representation of how scene points are expected to move in 3D. Rather than executing the predicted trajectory as an open-loop plan, GeomVLA extracts intermediate motion tokens from the trajectory denoiser and uses them to condition a 3D flow-based action denoiser through geometry-aware attention. GeomVLA achieves state-of-the-art performance on CALVIN, competitive performance on LIBERO and RoboTwin2.0, and outperforms strong baselines in real-world manipulation settings without robot-action pretraining. Extensive ablations show that future-motion reasoning alone is insufficient: the primary gains are associated with maintaining geometric consistency among scene representation, motion prediction, and robot actions throughout the perception-to-action pipeline.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 15 Sep 2026 · of version 1 of the paper; version 2 is newer · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. GeomVLA makes a solid case that 3D geometric consistency across perception, motion, and action improves VLA performance, with thorough ablations backing the claim. This line is wrong
4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the references say.
4 minor objections
The paper says, at §3.2, Eq. (2)
Pith says: minor objection 1 of 4
The motion-weighted loss uses weights wi,t that depend on a soft motion-saliency term. The main text describes the weight formula only briefly in the appendix. Consider adding a short summary (e.g., the motivation for upweighting moving points while retaining static structure) to the main exposition, as the loss design affects the quality of learned motion features.
The paper says, at Table 1, RoboTwin2.0 column
Pith says: minor objection 2 of 4
The footnote explains that the five-task comparison uses different training sets (GeomVLA trained on all 50 tasks, while GeomVLA w/o motion and the joint-angle variants are trained only on five tasks). This is transparent but easily missed. Moving the caveat to the main caption or adding a note in the main text would improve clarity.
show 2 more
The paper says, at §4.3, Real-World Experiments
Pith says: minor objection 3 of 4
The number of demonstration rollouts varies (50 for three tasks, 20 for five). A brief rationale for this split (e.g., task difficulty, data collection budget) would help readers assess the fairness of the comparison across tasks.
The paper says, at Appendix, Table 4
Pith says: minor objection 4 of 4
The trajectory denoiser hidden dimension for the real-world experiments is 128, notably smaller than for simulation (320–512). It would be helpful to comment on whether this was due to dataset size or computational constraints, and whether the final results might be sensitive to this architectural choice.
What would settle it
Independently evaluate the 3D Scene Trajectory Denoiser's pseudo-labels by comparing with ground-truth trajectories from a motion-capture system on a subset of real-world tasks (e.g., Insert marker, Stack blocks), then retrain the GeomVLA CALVIN model with the corrected labels and recompute the full ablation table (Table 2) to verify that the performance gain from geometric consistency persists under verified supervision.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 15 Sep 2026
- Reader holds up
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee accept: 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This is a well-executed paper that delivers on its central claim: maintaining geometric consistency among scene representation, motion prediction, and robot actions in a shared 3D coordinate frame is the primary source of improvement, not future-motion reasoning alone. The architecture is clean—a 3D Scene Trajectory Denoiser feeding intermediate motion tokens into a 3D action denoiser via geometry-aware attention—and the ablations (Table 2) are the strongest part of the paper. They isolate the effect of 3D motion reasoning, the choice of motion token (early-flow beats full trajectory or encoder-only), the importance of a common frame (arbitrary rigid preserves performance; mixing 2D motion with 3D action hurts), and rule out a parameter-count confound. The two-seed stability check on CALVIN and LIBERO is responsible.
The empirical results are competitive: SOTA on CALVIN, strong on LIBERO and RoboTwin2.0, and consistent real-world gains across eight tasks with 20–50 trials each. The real-world setup with ALOHA2 and multi-view Kinects is realistic. The inclusion of a joint-angle variant (GeomVLA-JA) that still outperforms the 2D counterpart is a nice extra validation that the benefit lives in the 3D representation, not in the action parameterization.
Soft spots are modest and mostly acknowledged. The trajectory denoiser relies on SpatialTrackerV2 pseudo-labels without independent verification, though the motion representation's utility is confirmed through instruction-sensitivity tests (Table 7) and the parameter-count control. Real-world results are on 20 trials per task; some tasks (rank by color, place in triangle) show low absolute success, which the authors honestly attribute to cross-task interference. No code is released (common but notable). The five-task RoboTwin2.0 comparison uses different training sets, but this is transparently stated and the central ablations on CALVIN and LIBERO are fully controlled.
The paper is clearly written and the claims are proportionate to the evidence. It deserves a serious referee and will be a useful reference for anyone building 3D-aware VLA models.
The strongest objection
After careful review, the central claim—that maintaining geometric consistency among scene representation, motion prediction, and robot actions is the primary source of improvement—is well-supported by the ablation study (Table 2). The pseudo-label reliance (SpatialTrackerV2, §3.2) is a common limitation, but the motion representation's utility is confirmed through multiple controls: instruction sensitivity (Table 7), performance degradation when motion is removed or represented in 2D (Tables 1–2), and the parameter-count control (random init, Appendix). The RoboTwin2.0 five-task comparison (Table 1) uses different training sets, but this is transparently stated and does not affect the central ablation results on CALVIN and LIBERO, which are consistent and reproducible across two seeds. No fundamental flaw, inconsistency, or hidden assumption that would overturn the central claim was found.
The reader's weakest_assumption (pseudo-label accuracy) is technically correct but does not constitute a load-bearing concern for the central claim. The paper provides converging evidence that the motion representation is meaningful and beneficial, independent of the absolute accuracy of the pseudo-labels: instruction sensitivity (Table 7), horizon impact (Table 6), the parameter-count control, and the systematic ablation pattern all support the claim that geometric consistency is the primary driver of improvement. The RoboTwin2.0 training-set confound is acknowledged and does not affect the core ablation results on CALVIN and LIBERO, which are consistent across seeds. Therefore, the present stress-test identifies no objection that would warrant a change to the ACCEPT verdict.
The referee's report
If the claimed results hold, GeomVLA makes a solid contribution to the VLA literature by demonstrating that explicit 3D geometric consistency across perception, intermediate motion reasoning, and action control yields clear gains over both 2D alternatives and motion-only baselines. The ablation study is well-designed and isolates the effect of shared geometry from other factors. The method achieves competitive or state-of-the-art results on multiple benchmarks without large-scale robot-action pretraining, and the real-world evaluation adds practical credibility. The paper is clearly written and the architectural choices are well motivated.
Pith's reading of the paper (a machine's summary, not the authors')
The paper argues that vision-language-action (VLA) models for robot manipulation benefit when scene perception, latent motion prediction, and action control all operate in the same metric 3D space. Most prior VLAs either work in 2D or decouple these components, losing geometric consistency. GeomVLA lifts pretrained VLM features into 3D scene tokens using depth and camera calibration, introduces a task-conditioned 3D Scene Trajectory Denoiser that predicts future point motion, and feeds intermediate motion features into a 3D flow-based action denoiser via geometry-aware attention. Experiments show that maintaining a shared robot-centric 3D frame throughout the pipeline yields larger gains than future-motion reasoning alone, across simulation benchmarks and real-world tabletop tasks.
What carries the argument. The central mechanism is the 3D Scene Trajectory Denoiser, a flow-matching module that predicts per-step 3D displacement increments for a grid of image points lifted into the robot base frame. Its intermediate decoder state (at initial noise, not the final trajectory) is pooled into motion tokens and fused with scene tokens via a geometry-aware cross-attention layer, which then conditions a separate 3D rectified-flow action denoiser that outputs end-effector poses. Positions, predicted displacements, and action tokens all reside in the same robot-centric metric space.
Core claim. GeomVLA demonstrates that explicitly linking scene representation, future point-motion prediction, and action generation within a single 3D coordinate frame improves language-conditioned robot manipulation performance. The key finding is that geometric consistency across these three stages—not just predicting future motion in any representation—is responsible for the gains. Ablations show that using 2D motion features with a 3D action denoiser hurts performance, and that applying the same arbitrary rigid transformation to all three components preserves most of the benefit, confirming the primacy of shared geometry.
The premise it rests on. The 3D Scene Trajectory Denoiser is trained using pseudo-labels from an external point-tracker (SpatialTrackerV2); if those labels are systematically inaccurate or biased, the learned motion representation and downstream action policy may be degraded.
If the paper is right (Pith's inferences, not the authors')
Future VLA architectures should embed perception, reasoning, and control in a common 3D coordinate system rather than treating them as separate modalities.
Intermediate motion features from an early stage of the trajectory denoiser are more effective than final denoised trajectories for conditioning the action policy.
The method's reliance on pseudo-labels from SpatialTrackerV2 suggests that high-quality 3D point tracking from video can serve as cheap supervision for motion prediction in robotics.
GeomVLA's architecture supports both Cartesian end-effector control and joint-angle control by feeding intermediate Cartesian action features into a downstream joint-angle denoiser, showing that the 3D geometric representation benefits multiple action spaces.
The performance gap between GeomVLA and the variant using 2D motion with 3D actions implies that aligning the geometric domain of intermediate reasoning with the action space is at least as important as the reasoning itself.
Since GeomVLA is trained only on benchmark-scale demonstrations without large-scale robot-action pretraining, the shared 3D design may offer a path to data efficiency that could be amplified with more diverse data.
The 3D Scene Trajectory Denoiser acts as a learned motion prior that is task-conditioned; this suggests that combining geometric motion priors with language instructions could be a general principle for long-horizon manipulation.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
We present GeomVLA, a Vision-Language-Action (VLA) model that unifies perception, latent scene motion prediction, and action generation within a shared robot-centric 3D coordinate frame. Our approach lifts pretrained VLM features into spatially grounded 3D scene tokens using depth and camera calibration, while retaining the semantic representations learned during VLM pretraining. We further introduce a 3D Scene Trajectory Denoiser, a task-conditioned module that learns a latent representation of how scene points are expected to move in 3D. Rather than executing the predicted trajectory as an open-loop plan, GeomVLA extracts intermediate motion tokens from the trajectory denoiser and uses them to condition a 3D flow-based action denoiser through geometry-aware attention. GeomVLA achieves state-of-the-art performance on CALVIN, competitive performance on LIBERO and RoboTwin2.0, and outperforms strong baselines in real-world manipulation settings without robot-action pretraining. Extensive ablations show that future-motion reasoning alone is insufficient: the primary gains are associated with maintaining geometric consistency among scene representation, motion prediction, and robot actions throughout the perception-to-action pipeline.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.95
“Extensive ablations show that future-motion reasoning alone is insufficient: the primary gains are associated with maintaining geometric consistency among scene representation, motion prediction, and robot actions throughout the perception-to-action pipeline.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.74 · human participants 0.04 · controlled comparison 0.89 · randomized assignment 0.01 · states quantitative result 0.08 · compares with prior results 0.97 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v2). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| We present GeomVLA, a Vision-Language-Action (VLA) model that unifies perception, latent scene motion prediction, and action generation within a shared robot-centric 3D coordinate frame. | contribution 0.99 | 0.93 | 0.18 | asserted 1.00 |
| Our approach lifts pretrained VLM features into spatially grounded 3D scene tokens using depth and camera calibration, while retaining the semantic representations learned during VLM pretraining. | method 0.97 | 0.90 | 0.04 | asserted 1.00 |
| We further introduce a 3D Scene Trajectory Denoiser, a task-conditioned module that learns a latent representation of how scene points are expected to move in 3D. | contribution 0.54 | 0.94 | 0.04 | asserted 1.00 |
| Rather than executing the predicted trajectory as an open-loop plan, GeomVLA extracts intermediate motion tokens from the trajectory denoiser and uses them to condition a 3D flow-based action denoiser through geometry-aware attention. | method 0.98 | 0.91 | 0.07 | asserted 1.00 |
| GeomVLA achieves state-of-the-art performance on CALVIN, competitive performance on LIBERO and RoboTwin2.0, and outperforms strong baselines in real-world manipulation settings without robot-action pretraining. | result 1.00 | 0.96 | 0.41 | asserted 1.00 |
| Extensive ablations show that future-motion reasoning alone is insufficient: the primary gains are associated with maintaining geometric consistency among scene representation, motion prediction, and robot actions throughout the perception-to-action pipeline. | interpretation 0.81 | 0.96 | 0.64 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.95
What was examined, and what was not
Sources read
- arXiv listing abstract v2
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,135 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment