2nd Place Solution to the HANDS 2026 Workshop Challenge-Dexterous Grasp Motion Track: Single-Shot Trajectory Warping for Grasp Motion Generation
Muneeb A. Khan, Woojin Kim, Shinwoo Kim, Muhammad Munsif, Binod Bhattarai, Seungryul Baek
cs.RO · cs.AI · cs.CV
Quick verdict
2nd-place challenge report: one demo + 12-D warp, one-step PPO; 94.61% easy (best), 57.18% hard; honest ablations, thin stats, demo sensitivity untested.
Summary
This workshop report describes the 2nd-place solution to the HANDS 2026 challenge (ECCV 2026) on dexterous grasp motion: generating reach-and-lift trajectories for the 12-DoF LinkerHand O6 in Isaac Gym on unseen Objaverse objects from randomized initial poses. Rather than a per-step policy (~70 sequential 12-D decisions), the method edits a single successful GraspM3 demonstration: a policy observes the object once (frozen DexRep PointNet embedding plus palm/object pose) and outputs a 12-D warp — wrist shift, wrist rotation, and finger offset — applied object-centrically with position-only alignment, a verbatim-copied lift phase, proportional finger rescaling, and a prepended reach segment, then replayed open-loop. The warp policy trains with one-step PPO (a contextual bandit, no temporal credit assignment) over all 4,824 training objects in parallel on three RTX 3090s (~16 hours, two stages with rotating trajectory indices). It scored 94.61% on the easy track — the highest of any submission — and 57.18% on the hard track (randomized mass/friction), placing second behind HandsDown (91.29/77.45). Local evaluation shows 90.98% on held-out trajectories with no measurable trajectory overfitting, though 4,487 of 4,502 validation objects were seen in training. Limitations: object-level generalization rests only on leaderboard numbers; sensitivity to the single bottle demonstration is unexamined; no seeds or confidence intervals; and the easy-hard gap is attributed to open-loop replay without testing a closed-loop variant.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 16 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
- A single 12-D warp of one successful GraspM3 demonstration, replayed open-loop, achieves 94.61% on the challenge easy track — the highest score of any submission.
- One-step PPO (contextual bandit) trained over all 4,824 objects in parallel scales performance: moving from 64 to 4,824 training objects adds +30.7 pp on easy and +17.5 pp on hard.
- Position-only object-centric alignment is essential: additionally aligning to object rotation drops no-warp replay success from 23.8% to 4.8%.
- The remaining easy-to-hard gap (57.18% hard, 60.4% retention) is structural — caused by open-loop replay, dynamics-invisible warping, and inherited grip force — and requires feedback to close.
Strengths
Effective problem reformulation with best-in-field easy-track result
Collapsing ~70 sequential 12-D decisions into a single contextual-bandit action (Sec. 2, Fig. 1) yields 94.61% on the easy track — the highest of all submissions (Table 1) — versus the 23.93% behavior-cloning baseline.
Quantified, honest design ablations
Position-only alignment (23.8% vs 4.8% no-warp replay with rotation alignment, Sec. 2.1), object scaling 64 to 4,824 objects (+30.7 pp easy / +17.5 pp hard, Sec. 3), and the learning-rate resume collapse and fix (80.3% to 45.6% within 24 iterations; restart at 1e-5 raises held-out success 86.81% to 90.98%, Sec. 2.4) are all reported with concrete numbers.
Rare evaluation candor
Careful low-level correctness work
Diagnostic value of the failure analysis
Weaknesses
Object-level generalization not measured locally
Single-demonstration dependence unexamined
No uncertainty quantification
Untested headline conclusion
Uncontrolled baseline comparison
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a 2nd-place solution report for the HANDS 2026 workshop challenge (co-located with ECCV 2026), Dexterous Grasp Motion track. The task: from randomized initial hand poses, produce physically plausible reach-and-lift trajectories for the 12-DoF LinkerHand O6 in Isaac Gym on private Objaverse test objects, scored on an easy track (training-like physics) and a hard track (mass ~U[20,200] g, friction ~U[1,20]). Central claims: (i) a single successful GraspM3 demonstration, edited by a one-shot 12-D warp policy and replayed open-loop, reaches 94.61% on the easy track — the highest of any submission — and 57.18% on the hard track (rank 2 overall, Table 1); (ii) one-step PPO over all 4,824 training objects in parallel is sufficient to train the warp; (iii) the residual easy-to-hard gap is structural, caused by open-loop replay and dynamics-invisible warping. Claims (i)-(ii) are well supported; (iii) is plausible but untested against even a minimal feedback baseline.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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 this paper about?
It is a 2nd-place technical report for the HANDS 2026 workshop challenge (at ECCV 2026) on dexterous grasp motion. The authors generate reach-and-lift trajectories for a 12-DoF robotic hand (LinkerHand O6) in the Isaac Gym simulator, scoring 94.61% on the easy track (the best of all submissions) and 57.18% on the hard track against unseen objects.
How does the method work?
Instead of a per-step policy, a small network looks at the object once and outputs a single 12-dimensional 'warp' (wrist shift, wrist rotation, finger offset) that edits one pre-recorded successful grasp demonstration. The edited trajectory — with the lift phase copied unchanged and fingers rescaled proportionally — is replayed open-loop. The warp policy is trained with one-step PPO across all 4,824 training objects in parallel in about 16 hours on three GPUs.
Why did it place 2nd despite the best easy-track score?
Overall ranking weighs the hard track, where object mass and friction are randomized. The open-loop replay cannot react if the object slips during the lift, so the method keeps only 60.4% of its easy-track performance (57.18% hard), while the winner HandsDown keeps 84.8% (77.45% hard).
Is code or data available?
The report does not announce a code release. It fully specifies the warp equations (Eqs. 2-6), network and PPO hyperparameters, and points to the public challenge repository and the GraspM3 dataset, so reimplementation is possible but nontrivial (Isaac Gym, three RTX 3090s, ~16 hours of training).
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 1-3 are represented only by captions in the extracted text; qualitative claims (Fig. 3 rollouts) cannot be verified.
- Leaderboard numbers (Table 1) are organizer-reported as of 1 Sept 2026 and not independently verifiable.
- A few training details are unstated (optimizer name, exact batch semantics of 'trajectory rotation'), slightly limiting reimplementation fidelity.
- Challenge rules and track configuration are known only via the referenced GitHub repository [3], not fully described in-text.
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.