Research map / Image generation / Fast diffusion sampling
EnergyAction: Unimanual to Bimanual Composition with Energy-Based Models
Mingchen Song, Xiang Deng, Jie Wei, Dongmei Jiang, Liqiang Nie, Weili Guan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spatial energy constraint (ablation) | EnergyAction · RLBench2 (20 demos) | average success rate: 76.6 -> 77.3 | Improved |
| add temporal energy constraint (ablation) | EnergyAction · RLBench2 (20 demos) | average success rate: 76.1 -> 77.3 | Improved |
| add temporal and spatial energy to composition (ablation) | E-Compose · RLBench2 (20 demos) | average success rate: 73.4 -> 77.3 | Improved |
| add energy-based composition and coordination (ablation) | Base policy · RLBench2 (20 demos) | average success rate: 35.1 -> 77.3 | Improved |
| add energy components at 100 demos (ablation) | Base policy · RLBench2 (100 demos) | average success rate: 50.5 -> 86.4 | Improved |
| add energy components at 5 demos (ablation) | Base policy · RLBench2 (5 demos) | average success rate: 15.2 -> 44.6 | Improved |
| use Flow Matching instead of DDIM for both arms (ablation) | EnergyAction · RLBench2 (20 demos) | average success rate: 75.3 -> 77.3 | Improved |
| use Flow Matching instead of DDPM for both arms (ablation) | EnergyAction · RLBench2 (20 demos) | average success rate: 76.2 -> 77.3 | No clear effect |
| use DDIM instead of Flow Matching for both arms at 100 demos (ablation) | EnergyAction · RLBench2 (100 demos) | average success rate: 86.4 -> 82.1 | Hurt |
| use adaptive denoising instead of fixed 5-step denoising (ablation) | EnergyAction · RLBench2 | mean denoising steps: 5 -> 1.79 and 1.27 | Improved |
| use early stop denoising instead of fixed 5-step denoising (ablation) | EnergyAction · RLBench2 | mean denoising steps: 5 -> 2.13 and 2.32 | Improved |
| add energy composition without any single-arm pretraining versus 3DFA (ablation) | EnergyAction · RLBench2 (20 demos), 0 single-arm tasks | success rate: 44.8 -> 52.3 | Improved |
Closest papers by meaning
- RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (ICLR 2025 · similarity 0.89)
- HumanoidGen: Data Generation for Bimanual Dexterous Manipulation via LLM Reasoning (NeurIPS 2025 · similarity 0.87)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.87)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.87)
- ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.87)
- PIVOT-R: Primitive-Driven Waypoint-Aware World Model for Robotic Manipulation (NeurIPS 2024 · similarity 0.87)
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.
