Venue: ICLR 2026 Category: Dexterous Manipulation — Grasping Trend tag: Data efficiency / sim-to-real Authors: Haoqi Yuan, Ziye Huang (equal), Ye Wang, Chuan Mao, Chaoyi Xu; corresponding Zongqing Lu. PKU + BeingBeyond + Renmin U.

flowchart LR
  Demo["Single successful grasp trajectory<br/>{q_t*, p_t*ee-obj}_{t=0..T_D}"] --> Editor[Demo Editor π · single-step MDP]
  Obs[Initial p_0^ee, p_0^obj, full point cloud c_0^obj<br/>PointNet 128-d → MLP 1024-1024-512-512] --> Editor
  Editor --> Edit[T^ee ∈ SE(3) Δxyz∈[-0.05,0.05]m<br/>Δrpy∈[-1.57,1.57]rad<br/>Δq^G ∈ [-1,1]rad]
  Edit --> Replay[Open-loop replay in IsaacGym<br/>40 sim steps · motion-planned approach]
  Replay --> Reward[r = 1[success] · 1[no collision]<br/>collision randomly disabled in 50% envs]
  Reward --> PPO[PPO · 7000 parallel envs · lr=3e-4]
  PPO --> Pol[Universal state-based policy]
  Pol --> Roll[35K rendered rollouts<br/>RGB + DR + ViT]
  Roll --> FM[Flow-matching action head<br/>GR00T-N1.5 architecture]
  FM --> Real[Sim-to-real on FR3 + Inspire Hand]

Universal dexterous grasping is a multi-task RL problem with high-dimensional action spaces (typically ~24 DoF for an arm + dexterous hand) and long horizons. Prior work mitigated exploration via complex reward shaping, curriculum learning (UniDexGrasp, UniDexGrasp++), iterative distillation (UniGraspTransformer), or two-stage residual RL (ResDex). These methods (i) often run on floating-wrist hands without arms, (ii) use privileged contact information, (iii) face a trade-off between collision penalties and task rewards, (iv) struggle with small/thin tabletop objects, and (v) require complex pipelines that are hard to extend to new embodiments.

DemoGrasp's reformulation: a single successful grasp trajectory encodes most of the transferable structure of universal grasping (approach center, squeeze, lift). Instead of exploring in the low-level robot action space, explore in the space of edits applied to that trajectory.

The demonstration D is a successful grasp trajectory of one specific object (collected by teleop or hard-coded), expressed in the initial object frame (world frame translated to the object's geometric center at t=0):

D = {(q_t^*hand, p_t^*ee-obj)}_t=0..TD

where q^*hand is hand joint targets and p^*ee-obj is the 6D end-effector pose in the initial object frame.

Two parameters: an SE(3) end-effector transformation T^ee and delta hand grasp pose Δq^G. Edited targets:

• p_t^*′ee-obj = T^ee p_t^*ee-obj for t ≤ T_lift; vertical lift Δz afterward.
• q_t^*′hand linearly interpolates from q₀^*hand to (q_Tlift^*hand + Δq^G) before T_lift; held thereafter.

Replaying D′ from a motion-planned starting pose works open-loop and already gives ~75% success on training objects (Table 8 row 1).

• State: initial end-effector 6D pose p₀^ee, initial object 6D pose p₀^obj, full object point cloud c₀^obj.
• Action: T^ee + Δq^G. End-effector translation Δxyz ∈ [-0.05, 0.05] m; rotation Δrpy ∈ [-1.57, 1.57] rad; hand DoFs Δq ∈ [-1, 1] rad. Rotations are quaternions in observation, Euler in action.
• Transition: replays edited demo for up to T = 40 simulation steps then terminates.
• Reward: r = 1[success] · 1[no collision during execution].

Strict no-collision can prevent grasping flat objects that need fingers to slip under them. The paper randomly disables robot–table collision detection in half of the parallel envs, so:

• Collision-free successful grasps: E[r] = 1
• Successful grasps with table contact: E[r] = 0.5
• Failures: E[r] = 0

This encourages collision-free grasps where possible while permitting minimal contact for hard objects.

• Algorithm: PPO (Schulman et al., 2017).
• Encoder: PointNet on full object point cloud → 128-d feature.
• Actor/critic MLP: [1024, 1024, 512, 512] with ELU activations; tanh output rescaled to action ranges.
• Hyperparameters (Table 12): parallel envs 7,000; learning rate 3e-4; PPO clip ε=0.2; episode length 1; demo-replay execution 40 steps; rollout steps/iteration 1; update epochs/iter 5; minibatches/epoch 4; init Gaussian σ=0.8; gradient clip 1.0.
• Compute: 24 hours on a single NVIDIA RTX 4090.

• 35,000 rendered trajectories collected with the trained RL policy; only successful ones kept.
• Architecture: GR00T-N1.5 (ViT encoder + flow-matching action head). Pretrained ViT fine-tuned; action head trained from scratch (no GR00T-N1.5 weights used). 100K iterations on 4× A800 GPUs, 16 hours.
• Domain randomization: RoboTwin's 300 background images for table textures; 100 FMD material images for object textures; 3 point lights with intensity/ambient ∈ [0.1, 0.8]; camera extrinsics noise [-0.02, 0.02]; table position noise [-0.05, 0.05] m.
• Depth augmentation: depth-range randomization, Gaussian blur+noise, 1% pixel dropout, blending with NYU-Depth-v2 frames at α=0.005 (HERMES recipe).
• Real hardware: 7-DoF Franka Research 3 (FR3) arm + 6-DoF Inspire Hand (6 active + 6 passive); 2× RealSense D435i cameras at diagonal viewpoints.

3,200 training objects; floating 6-DoF wrist + 18-DoF Shadow Hand. Spatial randomization 50 cm × 50 cm (baselines do not randomize position).

+5 state / +4 vision over best baseline; only 1% generalization gap between train and unseen — strong universality.

Trained on 175 objects (75 YCB + 100 DexGraspNet). Tested on 5 OOD datasets:

DemoGrasp matches RobustDexGrasp on ModelNet40 and exceeds on the other four datasets.

Six embodiments tested on the same 175-object training set + five test datasets:

• Inspire Hand (FR3+Inspire), Allegro (UR5+Allegro), DClaw (FR3+DClaw), floating Shadow, FR3+Shadow, Schunk SVH (UR5+Schunk), and FR3+Gripper (Panda parallel-jaw).
• All multi-fingered hands achieve >90% on the 175 training objects.
• Average 84.6% across all six unseen test datasets.
• Floating Shadow vs FR3+Shadow: difference of only 1.4% — adding a robot arm doesn't hurt.
• FR3+Gripper underperforms on EGAD/DGA (Panda's limited stroke struggles with wide objects).
• Per-dataset average ranges (across all six embodiments): EGAD 49.4–97.9%, DGA 30.7–79.3%, Omni6DPose 66.6–85.0%, DexGraspNet (Unseen Cat.) 81.1–97.5%, ModelNet40 61.2–80.1%, Visual Dexterity 83.5–99.1%.

110 unseen objects, 5 trials each, 50×50 cm position randomization:

95.3% on normal-sized objects. 68.3% on flat/thin (thickness < 1.5 cm), 76.7% on small (diameter < 3.5 cm). The paper claims this is the first universal sim-to-real policy to grasp small/thin tabletop objects without severe collisions.

Language-conditioned slightly outperforms unconditional (the paper attributes this to reduced action uncertainty for imitation learning). Robust to randomized backgrounds + lighting (82% under challenging scene appearance changes). Uses GR00T-N1.5's vision-language model frozen for instruction encoding.

Training BC on uniformly sampled successful rollouts gives multimodal/inconsistent data; RL converges to a unimodal optimal policy.

Measures contribution of each editing axis:

Adding wrist rotation (Δrpy) is the biggest win (+13 train); hand DoFs add a smaller +2 — using a dexterous hand as a single-DoF gripper already gets most of the way, but Δq enables more robust grasps (vase from the side using thumb+index+ring force closure).

Only 2.4% average gap between training on 175 objects and training on the test sets — universal grasping doesn't need millions of training objects with this method.

Even when the demo's open-loop replay is nearly useless (3.88% on a "big-object side grasp" demo), RL recovers >95% — DemoGrasp is robust to demo quality.

RGB beats depth on small/flat objects (sensor noise hides them in depth); two views beat one (richer 3D, less hand-object occlusion).

The paper does not have an explicit "Limitations" section. Implicit/honest gaps surfaced in the experiments:

• Flat/thin objects still hardest — 68.3% real-world, vs 95.3% on normal objects. The collision-free/with-contact reward design helps but the gap remains.
• Single-step open-loop replay depends on a successful motion plan from the home pose to the demo's first frame; cluttered start states could fail.
• Policy is closed-loop only on the vision side — the underlying RL action space is single-step on the editing parameters. For dynamic objects or major mid-grasp disturbances, the open-loop replay window may be short.
• Reset region is 50 cm × 50 cm tabletop — out-of-region or non-tabletop scenarios (deep bins, shelves, occluded objects) are not evaluated.
• No in-hand reorientation, tool use, bimanual — universal "pick and lift" only.

DemoGrasp is the cleanest current example that problem reformulation > scaling for dexterous skills.

• vs UniDexGrasp / UniDexGrasp++ / UniGraspTransformer: all explore in the full robot action space with complex curricula and reward shaping. DemoGrasp's compact 6+J editing space + single-step MDP + binary reward beats them with simpler training.
• vs ResDex / RobustDexGrasp: ResDex uses a two-stage residual RL framework; RobustDexGrasp focuses on robustness via large-scale data. DemoGrasp uses one demonstration and beats RobustDexGrasp on 4/5 OOD datasets with the Allegro Hand.
• vs DextrAH-G / DextrAH-RGB / Singh et al.: these are sim-to-real on a wide variety of objects but still fall short on small/thin tabletop objects. DemoGrasp is, per the paper, the first to grasp previously unseen small/thin objects in tabletop settings without severe collisions.
• vs ClutterDexGrasp: ClutterDexGrasp adds clutter; DemoGrasp adds clutter as an extension and still hits >80%.
• vs DexNDM: DexNDM tackles dexterous policies via a different angle (neural distance manifold). Both papers exemplify the 2026 trend that careful problem framing replaces brute-force scaling.
• vs D-REX: D-REX is a robot-experience scaling effort; DemoGrasp is on the data-efficiency end of the same trend.

The single-step MDP + demo-editing recipe is broadly transferrable to other manipulation skills (pouring, insertion) where a successful trajectory exists and only needs re-aiming to generalize.

• OpenReview: https://openreview.net/forum?id=Bf4FeuW0Mr
• PDF: https://openreview.net/pdf?id=Bf4FeuW0Mr
• Project page: https://research.beingbeyond.com/demograsp

• DexNDM
• ClutterDexGrasp
• D-REX

← Back to ICLR-2026