Research map / Reinforcement learning / Visual and representation RL
Efficient Residual Learning with Mixture-of-Experts for Universal Dexterous Grasping
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ResDex, combining residual RL with a mixture-of-experts of geometry-unaware base policies to learn a multi-task universal dexterous grasping policy across 3,200 DexGraspNet objects in about 12 GPU-hours.
Problem
Universal dexterous grasping across thousands of objects with RL suffers from multi-task gradient interference, complex curricula, long training, and poor generalization to unseen objects.
Approach
Train geometry-unaware base policies (proprioception and object position only) on cluster-center objects, then learn a hyper-policy that outputs a residual action plus MoE weights over the base policies' actions, using residual multi-task RL with a two-stage reward.
Key results
- Second-stage state-based ResDex (k=4) reaches 94.6% train success vs. UniDexGrasp++ 87.9% (first stage), with no generalization gap across unseen objects (94.4%, 95.4%).
- Vision-based ResDex (k=4) achieves 88.8% train success, 88.5% unseen objects in seen categories, 87.2% unseen categories, outperforming UniDexGrasp++ (85.4/79.6/76.7).
- Training takes about 20 min per base policy and 11 hours for the hyper-policy on a single RTX 4090 for state-based policies.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ResDex with four geometry-unaware experts and residual learning instead of UniDexGrasp++ | State-based policy · DexGraspNet train, stage 2 | success rate: 87.9 -> 94.6±1.6 | Helped |
| Use ResDex with four geometry-unaware experts and residual learning instead of UniDexGrasp++ | State-based policy · DexGraspNet unseen objects, seen categories, stage 2 | success rate: 84.3 -> 94.4±1.7 | Helped |
| Use ResDex with four geometry-unaware experts and residual learning instead of UniDexGrasp++ | State-based policy · DexGraspNet unseen categories, stage 2 | success rate: 83.1 -> 95.4±1.0 | Helped |
| Distill ResDex (k=4) into a vision-based policy instead of UniDexGrasp++ | Vision-based policy · DexGraspNet train | success rate: 85.4 -> 88.8 | Helped |
| Distill ResDex (k=4) into a vision-based policy instead of UniDexGrasp++ | Vision-based policy · DexGraspNet unseen objects, seen categories | success rate: 79.6 -> 88.5 | Helped |
| Distill ResDex (k=4) into a vision-based policy instead of UniDexGrasp++ | Vision-based policy · DexGraspNet unseen categories | success rate: 76.7 -> 87.2 | Helped |
| Add residual actions to the MoE hyper-policy (ablation) | ResDex multi-task policy · DexGraspNet train | success rate: MoE: 61.4, 71.1, 79.4, 80.3, 72.1, 81.6; MoE+Res: 83.2, 82.8, 88.1, 90.6, 87.6, 88.7 | Helped |
| Use geometry-unaware observations rather than full state observations or the full grasping proposal reward (ablation) | Single-object base policies · DexGraspNet training set generalization evaluation | success rate | Helped |
| Vary the number of geometry-unaware experts from k=1 to k=6 (ablation) | ResDex hyper-policy · DexGraspNet unseen categories | grasp-pose D: 223.6, 174.5, 194.3, 176.3, 204.6, 176.1 | Mixed |
Surprising findings
- Geometry-unaware base policies trained on a single object generalize better than policies with full observations or the full grasping proposal reward.
- Adding residual actions to MoE improved training success substantially (e.g., k=1: 61.4% to 83.2%), while the k=5 configuration performed worse than k=4 in the first stage.
Limitations
- Not yet extended to functional grasping tasks.
- Vision-based policy not deployed on real hardware (sim-to-real gap unaddressed).
Experimental setup
Datasets: DexGraspNet
Benchmarks: DexGraspNet
Metrics: success rate D (grasping pose quality, lower is better)
Baselines: UniDexGrasp UniDexGrasp++ MoE without residual actions
Builds on: residual policy learning UniDexGrasp UniDexGrasp++ DAgger PPO
Closest papers by meaning
- UniGraspTransformer: Simplified Policy Distillation for Scalable Dexterous Robotic Grasping (CVPR 2025 · similarity 0.89)
- ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.86)
- Cross-Embodiment Dexterous Grasping with Reinforcement Learning (ICLR 2025 · similarity 0.86)
- DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness (CVPR 2025 · similarity 0.86)
- Scaffolding Dexterous Manipulation with Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.