Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/21/2026, 4:29:29 AM
Summary
This paper presents a second-place solution to the HANDS 2026 Workshop Challenge (Dexterous Grasp Motion track) using a single-shot trajectory warping method. The approach edits a single successful GraspM3 demonstration via a 12-dimensional warp output by a policy that observes the object once, rather than generating motion step-by-step. Trained with one-step PPO over 4,824 objects in parallel using Isaac Gym, the method achieved 94.61% success on the easy track and 57.18% on the hard track.
Entities (8)
Relation Signals (8)
Proposed Method → achievedrank → 2nd Place
confidence 99% · 2nd place solution to the HANDS 2026 workshop challenge
HANDS 2026 Workshop Challenge → hastrack → Dexterous Grasp Motion Track
confidence 99% · HANDS 2026 workshop challenge (Dexterous Grasp Motion track)
Proposed Method → achievedsuccessrate → 57.18%
confidence 98% · and 57.18% on the hard track
Proposed Method → achievedsuccessrate → 94.61%
confidence 98% · achieved success rates of 94.61% on the easy track
Proposed Method → trainedwith → PPO
confidence 98% · train the warp policy with one-step PPO
LinkerHand O6 → usedin → HANDS 2026 Workshop Challenge
confidence 97% · address grasp motion generation for the 12-DoF LinkerHand O6
Proposed Method → simulatedin → Isaac Gym
confidence 96% · replaying them in Isaac Gym
Proposed Method → uses → GraspM3
confidence 95% · editing a single successful GraspM3 demonstration
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:This report describes our 2nd place solution to the HANDS 2026 workshop challenge (Dexterous Grasp Motion track) in conjunction with ECCV 2026. In this challenge, we address grasp motion generation for the 12-DoF LinkerHand O6, aiming to produce physically plausible reach-and-lift trajectories for unseen objects from randomized initial hand poses in simulation. This task is particularly challenging because each grasp requires a per-step policy to make approximately $70$ twelve-dimensional decisions, with errors accumulating over time, while test objects and physical dynamics may differ from those encountered during training. To address these challenges, we propose editing a single successful GraspM3 demonstration instead of generating the motion step by step: a policy observes the object once and outputs a 12-D warp of the demonstration, which is then replayed open-loop. Moreover, we train the warp policy with one-step PPO over all $4{,}824$ training objects in parallel. As a result, our method achieved success rates of $94.61\%$ on the easy track, the highest of all submissions, and $57.18\%$ on the hard track of the private test set.
Tags
Links
- Source: https://arxiv.org/abs/2609.21511v1
- Canonical: https://arxiv.org/abs/2609.21511v1
Trouble viewing inline? Open PDF directly →
Full Text
19,175 characters extracted from source content.
Expand or collapse full text
2nd Place Solution to the HANDS 2026 Workshop Challenge-Dexterous Grasp Motion Track: Single-Shot Trajectory Warping for Grasp Motion Generation Muneeb A. Khan Affiliation: UNIST, South Korea Woojin Kim Affiliation: UNIST, South Korea Shinwoo Kim Affiliation: UNIST, South Korea Muhammad Munsif Affiliation: UNIST, South Korea Binod Bhattarai Affiliation: University of Aberdeen, UK Affiliation: Fogsphere (Redev AI Ltd, UK) Seungryul Baek Affiliation: UNIST, South Korea Abstract This report describes our 2nd place solution to the HANDS 2026 workshop challenge (Dexterous Grasp Motion track) in conjunction with ECCV 2026. In this challenge, we address grasp motion generation for the 12-DoF LinkerHand O6, aiming to produce physically plausible reach-and-lift trajectories for unseen objects from randomized initial hand poses in simulation. This task is particularly challenging because each grasp requires a per-step policy to make approximately 7070 twelve-dimensional decisions, with errors accumulating over time, while test objects and physical dynamics may differ from those encountered during training. To address these challenges, we propose editing a single successful GraspM3 demonstration instead of generating the motion step by step: a policy observes the object once and outputs a 12-D warp of the demonstration, which is then replayed open-loop. Moreover, we train the warp policy with one-step PPO over all 4,8244,824 training objects in parallel. As a result, our method achieved success rates of 94.61%94.61\% on the easy track, the highest of all submissions, and 57.18%57.18\% on the hard track of the private test set. 1 Introduction Dexterous grasping is a prerequisite for general-purpose manipulation. A multi-fingered hand covers a far wider range of objects and grasp types than a parallel-jaw gripper, but the added dexterity brings a high-dimensional action space and contact-rich dynamics that are hard to model [4]. Two developments have made data-driven approaches practical, including large synthetic corpora [2] that pair objects with hand poses or complete grasp motions, and massively parallel simulators [6] that amortize exploration over thousands of objects at once. Combining them is nontrivial, because a corpus prescribes what the hand should do for one object instance, whereas a deployed policy must decide what to do for an unseen object from an arbitrary starting pose. The Dexterous Grasp Motion track [3] is built around this gap. It builds on GraspM3 [4, 9], retargeted from the Shadow Hand to the LinkerHand O6 consist of six actuated finger joints on a free-floating wrist, so that every control step is a 12-D command (wrist translation, wrist rotation, six finger targets). Training data covers 5,0485,048 ShapeNet [1] objects with ∼110 110k trajectories; replaying them in Isaac Gym [6] and keeping the successful ones leaves 4,8244,824 objects and 52,71852,718 trajectories. Episodes start 1515–2020 cm from a randomly oriented object and succeed on a 3030 cm lift. Submissions are scored on private Objaverse [2] objects on an easy track (mass 2020 g, friction 2020, as in training) and a hard track (mass ∼[20,200] [20,200] g, friction ∼[1,20] [1,20]). The official baseline is behavior cloning on DexRep [5] features (23.93%23.93\%). Per-step policies face significant hurdles in this context, requiring approximately 70 twelve-dimensional decisions per grasp from a 1,7571,757-D observation, with compounding error. DemoGrasp [10] showed that neither is strictly necessary for the geometric phase of grasping: one successful demonstration, edited by a policy that decides where the hand goes and how far it closes, generalizes across objects when the editor is trained as a single-step MDP. We port this formulation to the O6 task and GraspM3 data. Our contributions are (i) an object-centric warp of one demonstration with position-only alignment, verbatim lift preservation, and proportional finger rescaling (Sec. 2.2); (i) one-step PPO over all 4,8244,824 training objects in parallel with trajectory rotation (Sec. 2.4); and (i) an analysis of the easy-to-hard gap that an open-loop warp cannot close (Sec. 3). P∈ℝ512×3P ^512× 3 palm pose, object pose PointNet encoder Warp policy πθ(a|o) _θ(a\,|\,o) Warp τ by (Δp,Δθ,Δq)( p, θ, q) Eq. 4–6 single demonstration τ Π :reach∥\|track∥\|settle Isaac Gym rollout r=[lift≥0.3m]r=1[\,lift≥ 0.3\,m\,] zzaaτ -step PPO update of θ perceive once (t=0t=0) deterministic warping open-loop replay Figure 1: Overview of the trajectory warping pipeline. A frozen () PointNet embeds the object point cloud P; with the palm and object poses, and the embedding z is passed to the warp policy πθ _θ once per episode. 2 Method Figure 1 shows the pipeline. An episode provides an object o (initial pose and point cloud) and a randomized initial hand pose. The policy πθ _θ observes the scene once and outputs a∈[−1,1]12a∈[-1,1]^12; a warps a single demonstration τ into a full trajectory, which is replayed open-loop, and the episode returns r=[lift≥0.3m]r=1[lift≥ 0.3\,m]. Learning is the contextual bandit maxθo∼,a∼πθ(⋅|o)[r(o,a)], _θ\ E_o ,\,a _θ(·|o) [r(o,a) ], (1) with no credit assignment across time. 2.1 Object-centric demonstration We utilize a successful GraspM3 bottle grasp demonstration, verified via a single replay in Isaac Gym to ensure successful object lifting (Fig. 2). We record the commanded actions; replaying the resulting physical states would re-introduce a one-step execution lag: τ=utt=1Tτ=\u_t\_t=1^T, ut=(xt,rt,qt)∈ℝ3×ℝ3×ℝ6u_t=(x_t,r_t,q_t) ^3\!×\!R^3\!×\!R^6, T=69T=69, with lift onset tℓ=34t_ =34 (first frame with object rise >2>2 cm). The wrist command maps to the palm pose as pt=xt+cp_t=x_t+c, Rt=Rx(ϕt)Ry(θt)Rz(ψt)R_t=R_x( _t)R_y( _t)R_z( _t) (intrinsic XYZ), with a constant offset c measured at reset. We verified this convention numerically; Isaac Gym’s extrinsic Euler helper is off by 55 cm and would corrupt every warped frame. The demonstration is expressed relative to its initial object position only, p~t=pt−o0demo,R~t=Rt. p_t=p_t-o_0^demo, R_t=R_t. (2) Aligning to the object rotation as well drops no-warp replay success from 23.8%23.8\% to 4.8%4.8\%: GraspM3 object rotations are near-random per trial (‖ΔR‖F∈[0.77,2.83]\| R\|_F∈[0.77,2.83] within one object) and carry no approach information. Figure 2: A trajectory demonstration and sampled warps. (a) Palm path of τ (black; solid: approach and grasp, dashed: lift) and three warps (colors): Δθ θ rotates the approach about the object, Δp p shifts it, and the lift is copied verbatim (Eq. 5). (b) Replay in simulation: fingers close over steps 2525–3131, lift onset tℓ=34t_ =34, 0.300.30 m rise by t=69t=69. 2.2 Demonstration warping Following the DemoGrasp scales, the action is split into a wrist shift, a wrist rotation, and a finger offset, Δp=spa1:3,Δθ=sθa4:6,Δq=sqa7:12, p=s_p\,a_1:3, θ=s_θ\,a_4:6, q=s_q\,a_7:12, (3) with (sp,sθ,sq)=(0.05m, 1.57rad, 1.0rad)(s_p,s_θ,s_q)=(0.05\,m,\,1.57\,rad,\,1.0\,rad) and RΔ=Rx(Δθ1)Ry(Δθ2)Rz(Δθ3)R_ =R_x( _1)R_y( _2)R_z( _3). Wrist. Pre-lift frames are rigidly warped in the object-centric frame, rotating the approach about the object and shifting it: p^t=RΔp~t+Δp,R^t=RΔR~t(t<tℓ). p_t=R_ \, p_t+ p, R_t=R_ \, R_t (t<t_ ). (4) Lift frames copy the demonstration’s relative displacement onto the warped grasp pose, so the vertical motion that made the demonstration succeed is never rotated or scaled: p^t=p^tℓ−1+(p~t−p~tℓ−1)(t≥tℓ). p_t= p_t_ -1+( p_t- p_t_ -1) (t≥ t_ ). (5) The path returns to the world frame of the new episode (object at o) and to the action encoding: pt′=p^t+op _t= p_t+o, xt′=pt′−cx _t=p _t-c, rt′=EulerXYZ−1(R^t)r _t=Euler^-1_XYZ( R_t). Fingers. Let q1q_1 be the open pre-grasp pose and qtgq_t_g, tg=tℓ−1t_g=t_ -1, the closed grasp. The target grasp, q⋆=clip(qtg+Δq,qmin,qmax)q =clip(q_t_g+ q,\,q_ ,\,q_ ) is achieved by scaling the closing motion on a per-joint basis: λ=(q⋆−q1)⊘(qtg−q1),qt′=q1+λ⊙(qt−q1)t<tgq⋆t≥tg,λ=(q -q_1) (q_t_g-q_1),q _t= casesq_1+λ (q_t-q_1)&t<t_g\\ q &t≥ t_g, cases (6) so the closure morphs continuously into q⋆q with unchanged timing and the fingers keep squeezing through the lift. Trajectory synthesis. The randomized reset pose (x0,R0)(x_0,R_0) is generally far from the warped first frame (x1′,R1′)(x _1,R _1). We prepend a reach segment (LERP on position, SLERP on orientation) of n=max(⌈‖x1′−x0‖/0.01⌉,⌈∠(R0,R1′)/0.1⌉)n= \! ( \|x _1-x_0\|/0.01 ,\ (R_0,R _1)/0.1 ) steps, i.e. at most 11 cm and 0.10.1 rad per step, with fingers held at q1′q _1. The plan Π=[reach1:Tr∥track1:T∥settle1:10] =[reach_1:T_r\,\|\,track_1:T\,\|\,settle_1:10] is executed by sending each frame as an absolute 12-D target to the task’s PD controller; the policy is never called again. 2.3 Warp policy and one-step PPO The observation is the palm pose (7), the object pose (7), and a fixed-index subsample of 512512 points of the task’s object-local cloud, encoded by the ShapeNet PointNet [7] auto-encoder shipped with the DexRep baseline (frozen); only a linear projection to z∈ℝ128z ^128 is trained. Actor and critic are separate MLPs [1024,1024,512,512][1024,1024,512,512] with ELU and orthogonal initialization. The policy is a tanh-squashed Gaussian, a=tanh(a~)a= ( a), a~∼(μθ(o),diagσ2) a ( _θ(o),diag\,σ^2), with σ learned (initialized at 0.80.8) and the tanh Jacobian correction in logπθ _θ. With horizon one there is no bootstrapping: the return is r, Vϕ(o)V_φ(o) is a variance-reducing baseline, and A A is r−Vϕ(o)r-V_φ(o) standardized over the batch. We minimize the clipped PPO [8] objective ℒ= =\; [max(−A^ρθ,−A^clip(ρθ,1−ε,1+ε))] \! [ \! (- A _θ,\,- A\,clip( _θ,1- ,1+ ) ) ] +cv[(r−Vϕ(o))2], +c_v\,E\! [(r-V_φ(o))^2 ], (7) where ρθ=πθ(a|o)/πθold(a|o) _θ= _θ(a|o)/ _ _old(a|o) denotes the policy ratio. We use a clipping threshold of ε=0.2 =0.2 and a value coefficient of cv=2c_v=2. To ensure training stability, we clip gradients to a maximum norm of 1.01.0. Furthermore, we employ an adaptive learning rate strategy based on the Kullback-Leibler (KL) divergence. The learning rate is scaled by a factor of 1.51.5 to maintain a target KL divergence of 0.0160.016, and is strictly bounded within the interval [10−5,10−2][10^-5,10^-2]. 2.4 Large-scale training Each iteration runs all 4,8244,824 training objects in parallel, one environment per object, sharded over three GPUs (11.511.5 s per iteration); gradients, the KL statistic, and the advantage statistics are all-reduced, so the run equals a single 4,8244,824-environment batch. Because object scale is fixed at simulation build time, we rebuild the simulation every K iterations with the next trajectory index, covering 51,65951,659 of 52,71852,718 trajectories (98%98\%). The training process converged over 4,0004,000 iterations (approximately 16 hours). The initial 2,0002,000 iterations utilized trajectory indices 00–33, rotated every 500500 steps at an initial learning rate of 3×10−43× 10^-4, then 2,0002,000 with indices 00–1919 rotated every 100100. Resuming the converged stage-1 policy at the default learning rate collapsed it from 80.3%80.3\% to 45.6%45.6\% training success within 2424 iterations; restarting at the rate the KL controller had settled on (10−510^-5) raised held-out success monotonically from 86.81%86.81\% to 90.98%90.98\%. 3 Experiments Setup. We use three RTX 3090 GPUs. Local evaluation is deterministic (mean action, one rollout per object) in chunks of 1,5001,500 objects, so peak memory is independent of object count; leaderboard numbers come from the organizers. Per the rules, our entrypoint changes four settings without editing the task: full 12-D absolute targets every step, the pose observation in place of per-step DexRep, vision enabled, and start-position canonicalization disabled. Initial poses and the success test are unchanged. Training uses only GraspM3 objects; the sole pretrained component is the baseline’s frozen PointNet. Team (submission) Easy [%] Hard [%] Rank HandsDown (2) 91.29 77.45 1 UVLL HandDex (ours) 94.61 57.18 2 J-Team 87.20 43.88 3 Perception 68.80 23.98 4 OverGrasp 34.37 8.08 5 HCBHand 32.23 8.77 6 BC baseline (DexRep-TCN) [5] 23.93† – – Table 1: Leaderboard (private Objaverse test set, as of 1 Sept 2026): each team’s highest-ranked submission, ranked by team. †Reported by the organizers on their demo split. Checkpoint Evaluation split Objects Success [%] Final (stage 2) held-out traj. (valid) 4,502 90.98 Final (stage 2) training traj. (train) 4,824 90.78 Stage 1 held-out traj. (valid) 4,502 86.81 64-object model 64 held-out objects 64 57.27‡ Table 2: Local evaluation (deterministic, one rollout per object). The valid split shares no grasp sequence with train, but 4,4874,487 of its 4,5024,502 objects appear in training. ‡Mean over 10 rounds on held-out objects; not comparable to the rows above. Results. Table 1 compares teams. Our final submission scores 94.61%94.61\% on the easy track, the highest easy-track number of any submission, and 57.18%57.18\% on the hard track; HandsDown wins overall with 91.29/77.4591.29/77.45. Scaling the training set from 6464 objects (our second submission, 63.91/39.7063.91/39.70) to all 4,8244,824 added +30.7+30.7 p (easy) and +17.5+17.5 p (hard). Table 2 reports the local evaluation: held-out trajectories score 0.20.2 p higher than training trajectories, i.e. there is no measurable overfitting to the trained trajectories, and stage 1 alone reaches 86.81%86.81\%, so object coverage rather than iteration count is the main lever. Qualitative results and ablations. Figure 3 shows rollouts on a bottle and a toy animal: the warp places the palm on the surface, closes to a stable grasp, and the preserved lift completes the task. Failures follow one pattern: the object slips during the lift, and the plan proceeds without it. Three design choices had the largest effect: position-only alignment (23.8%23.8\% vs. 4.8%4.8\% no-warp replay on 6464 objects), object coverage (64→4,82464→ 4,824 objects: 63.9/39.7→94.6/57.263.9/39.7→ 94.6/57.2), and the learning rate on resumption (3×10−43× 10^-4 collapses a converged policy; 10−510^-5 improves it monotonically). Easy versus hard track. We retain 60.4%60.4\% of our easy-track success on the hard track; HandsDown retains 84.8%84.8\%. The gap is structural. (i) The plan is fixed before first contact, so an object that slides during the lift cannot be re-grasped. (i) Mass and friction are invisible to the warp, which sees only geometry, and training used easy-track physics only. (i) Grip force is inherited from the demonstration: finger targets are PD-tracked positions, and Δq q changes the shape of the grasp, not how hard it holds under ten times the load. The single decision therefore solves the geometric problem, where and how to grasp, while the remaining gap is dynamic and requires feedback. Figure 3: Qualitative evaluation of the final policy in simulation. 4 Conclusion We ported DemoGrasp’s single-demonstration, one-step RL formulation to the LinkerHand O6 and scaled it to every GraspM3 training object. One demonstration and one 12-D decision per episode reach 94.61%94.61\% on unseen Objaverse objects, the highest easy-track score in the challenge, though only 57.18%57.18\% under randomized mass and friction: the single decision solves the geometric problem, while closing the dynamic gap requires feedback. References [1] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, J. Xiao, L. Yi, and F. Yu (2015) ShapeNet: an information-rich 3D model repository. arXiv preprint arXiv:1512.03012. Cited by: §1. [2] M. Deitke, D. Schwenk, J. Salvador, L. Weihs, O. Michel, E. VanderBilt, L. Schmidt, K. Ehsani, A. Kembhavi, and A. Farhadi (2022) Objaverse: a universe of annotated 3D objects. arXiv preprint arXiv:2212.08051. Cited by: §1. [3] HANDS Workshop Organizers (2026) Dexterous grasp motion challenge 2026 (hands workshop at eccv 2026). Note: https://github.com/DexGraspMotionChallenge/DexGraspMotionChallenge2026 Cited by: §1. [4] H. Li, Q. Ye, Y. Huo, Q. Liu, S. Jiang, T. Zhou, X. Li, Y. Zhou, and J. Chen (2024) TPGP: temporal-parametric optimization with deep grasp prior for dexterous motion planning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), p. 18106–18112. Cited by: §1. [5] Q. Liu, Y. Cui, Q. Ye, Z. Sun, H. Li, G. Li, L. Shao, and J. Chen (2023) DexRepNet: learning dexterous robotic grasping network with geometric and spatial hand-object representations. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), p. 3153–3160. Cited by: §1, Table 1. [6] V. Makoviychuk, L. Wawrzyniak, Y. Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State (2021) Isaac Gym: high performance GPU-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470. Cited by: §1. [7] C. R. Qi, H. Su, K. Mo, and L. J. Guibas (2016) PointNet: deep learning on point sets for 3D classification and segmentation. arXiv preprint arXiv:1612.00593. Cited by: §2.3. [8] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Cited by: §2.3. [9] Q. Ye, H. Li, Q. Liu, S. Jiang, T. Zhou, Y. Huo, and J. Chen (2025) Contact2Motion: contact guided dexterous grasp motion generation with synergy embedded optimization. The International Journal of Robotics Research, p. 02783649251364392. Cited by: §1. [10] H. Yuan, Z. Huang, Y. Wang, C. Mao, C. Xu, and Z. Lu (2025) DemoGrasp: universal dexterous grasping from a single demonstration. arXiv preprint arXiv:2509.22149. Cited by: §1.