ProxiDex: Learning Dynamics-Guided Proximity Policy for Dexterous Manipulation
Yushan Bai, Boyu Zheng, Zhiyang Mao, Hongzheng Sun, Yuchuang Tong, En Li, Zhengtao Zhang
cs.RO · cs.AI
Quick verdict
Claims proximity dynamics lift dexterous manipulation success (sim 83.9% vs CordViP 81.0%, real 72.6% vs 62.9%); convincing, yet no significance tests or code.
Summary
ProxiDex targets multi-finger dexterous manipulation under partial observability, where the robot hand occludes contact regions and tactile sensing is costly. The framework reconstructs occlusion-free hand-object interaction point clouds (Grounded-SAM segmentation, SAM3D mesh reconstruction with metric-scale calibration, FoundationPose++ pose tracking) and converts hand-object distances into region-level proximity cues, used both as VR teleoperation feedback and as policy input. Training is two-stage: stage 1 pretrains a proximity VAE plus a DiT-based forward dynamics model (FDM), an inverse differential module (IDM) decoding proximity latents from latent changes, and a trajectory-phase head; stage 2 trains a Transformer diffusion policy with causal proximity masking and phase-adaptive gating. At inference, dynamics-consistency checking switches to an autoregressive 'dynamics-dream' mode when observations deviate from FDM predictions. Across 10 simulation tasks (Adroit, DexArt, IsaacLab), ProxiDex averages 83.9% success versus CordViP 81.0% and AFRO 76.9%; on real-robot tasks (Realman RM75B with 6-DoF and 20-DoF hands), 72.6% overall versus CordViP 62.9%, including unseen-object and perturbation settings. Ablations attribute roughly 7.5 points to interaction point clouds and 4.0 to FDM. Occlusion recovery reaches 80%/72% under 2-s/4-s occlusions but drops to 35% at 6 s. Limitations: dependence on object reconstruction and pose tracking (errors grow sharply beyond 0.8 m/s), short-term-only occlusion recovery, no force sensing, no code release, and no statistical significance testing.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Region-level hand-object proximity computed from reconstructed interaction point clouds is a hardware-agnostic contact representation that raises manipulation success (+2.9 points over CordViP in simulation, +9.7 on real robots).
- A coupled forward-inverse latent dynamics model (DiT FDM with Hdyn=2 plus an IDM aligning latent differences to VAE proximity latents) makes observation latents proximity-sensitive and enables autoregressive takeover during occlusion (80%/72% recovery at 2 s/4 s, degrading to 35% at 6 s).
- Trajectory-phase gating of proximity tokens plus dynamics-consistency-checked inference stabilizes control under perturbations and occlusion (67/90 vs CordViP's 56/90 perturbation trials).
- Geometric proximity is competitive with real tactile arrays for pre-grasp and contact establishment, though it cannot measure contact forces (Fig. 9).
- About 30 demonstrations recover roughly 94.7% of the 100-demonstration performance across six benchmark tasks (App. D.3).
Strengths
End-to-end system evaluated in both simulation and on real hardware
Secs. 4-5 cover data pipeline, dynamics model, policy, and inference in one coherent loop; real-robot evaluation spans 3 tabletop tasks plus 2 contact-rich tasks (Twist 10/20, Flip 11/20) on two different dexterous hands (6-DoF CasBot-P1L, 20-DoF Wuji Hand V1).
Substantial real-world evaluation protocol with per-setting trial counts
Table 2 reports 310 total trials across in-distribution (20 trials/task), unseen objects (40), and perturbation (30) settings; percentages are internally consistent (225/310 = 72.6%).
Granular ablations that support component attribution
Honest, quantified failure analyses
Unusually complete implementation detail
Weaknesses
No statistical significance testing anywhere
Real-robot robustness margins may be within binomial noise
'Hardware-agnostic' representation depends on an object model and a tracker
Qualitative tactile comparison and heuristic threshold
No code or data release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This review covers the full extracted text including Appendices A-E; figures were not analyzable, so figure-based findings rely on numbers quoted in the body. ProxiDex (arXiv 2609.16586, CoRL 2026 submission) proposes hand-object geometric proximity — computed from reconstructed interaction point clouds rather than tactile hardware — as an explicit, evolving interaction state for multi-finger manipulation. Three contributions: (1) a VR teleoperation pipeline with immersive proximity feedback replacing tactile instrumentation; (2) a coupled forward-inverse latent dynamics model (a DiT-based FDM predicting future observation latents, plus an IDM decoding proximity latents from latent differences); (3) a trajectory-adaptive Transformer diffusion policy with phase gating and dynamics-consistency-guided inference ('dynamics-dream' takeover). Headline claims: 83.9% average success over 10 simulation tasks (vs CordViP 81.0%) and 72.6% overall on real robots (vs CordViP 62.9%), with robustness under unseen objects, perturbations, and occlusions.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ProxiDex?
ProxiDex is a robot-learning framework for multi-finger dexterous manipulation. Instead of tactile sensors, it uses hand-object geometric proximity — computed from reconstructed interaction point clouds (object mesh via SAM3D plus pose tracking) — as a contact representation. It trains a two-stage policy (latent dynamics model plus diffusion policy) and provides proximity feedback to human operators during VR teleoperation for collecting training demonstrations.
How does ProxiDex handle visual occlusion of the hand?
At inference, a forward dynamics model predicts the next observation latent and compares it with the real encoded observation. When the deviation exceeds a threshold, the system assumes unreliable vision and switches to a 'dynamics-dream' mode, generating actions from autoregressively predicted latents. Recovery is 80%/72% after 2-s/4-s occlusions but drops to 35% at 6 s; the reliable open-loop horizon is roughly 3.5 seconds.
Does ProxiDex replace tactile sensing?
Not fully. Proximity captures hand-object distances and their spatial distribution before and around contact, and the paper reports competitive task performance versus a 2x4 tactile array per fingertip — but only qualitatively. Proximity cannot measure contact forces or fingertip force distribution, which the authors list as a limitation.
What results does ProxiDex report?
In simulation across 10 tasks (Adroit, DexArt, custom IsaacLab), it averages 83.9% success versus 81.0% for CordViP, 76.9% for AFRO, and 71.8% for DP3. On real robots (three tabletop tasks plus two contact-rich tasks, 310 trials), it reaches 72.6% overall versus 62.9% for CordViP, including unseen-object and perturbation conditions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Figs. 4-9, 11-22) are not analyzable from the extracted text; figure-based claims rely on in-text numbers quoted in captions or body.
- The number of evaluation episodes/seeds for simulation is not extractable from the text; statistical judgments there rest on reported stds only.
- arXiv identifier and CoRL 2026 venue are taken from the document header and were not independently verified.
- Minor typographic noise from PDF extraction (e.g., 'DexArt' task lists, subscripts) was normalized during review; table values were cross-checked for internal consistency.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.