Research map / Reinforcement learning / Visual and representation RL
Demystifying Action Space Design for Robotic Manipulation Policies
Yuchun Feng, Jinliang Zheng, Zhihao Wang, Dongxiu Liu, Jianxiong Li, Jiangmiao Pang, Tai Wang, Xianyuan Zhan
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 |
|---|---|---|---|
| predict delta actions instead of absolute actions (ablation) | EE-space ACT policy · Single Arm AgileX, average | progress score: 69.0 -> 89.6 | Improved |
| predict delta actions instead of absolute actions (ablation) | Joint-space DP policy · Single Arm AgileX, average | progress score: 85.0 -> 95.9 | Improved |
| predict delta actions instead of absolute actions (ablation) | Joint-space ACT policy · RoboTwin 2.0 | success rate: 40.0 -> 46.3 | Improved |
| predict delta actions instead of absolute actions (ablation) | Joint-space DP policy · Bimanual AgileX Bowl | progress score: 74.7 -> 74.6 | No clear effect |
| use chunk-wise delta instead of step-wise delta (ablation) | regression policy on single-arm AgileX · Touch Cube, Pick Up Cup, Pick and Place Cup | progress score: gap upwards of 10% on average | Improved |
| use shorter execution horizon for delta and longer for absolute actions (ablation) | action chunking · Single Arm AgileX, execution horizon 15 to 60 | progress score | Improved |
| use joint space instead of task (EE) space with flow-matching policy (ablation) | DP policy · Single Arm AgileX, average | progress score: 91.4 -> 95.9 (delta) | Improved |
| use joint space instead of task (EE) space with regression policy (ablation) | ACT policy, delta actions · Single Arm AgileX, average | progress score: 89.6 -> 88.0 | No clear effect |
| increase training epochs and demonstrations (ablation) | joint-space policies · Single Arm AgileX scaling | progress score | Improved |
| use task-space instead of joint-space actions (ablation) | cross-embodiment and pi0 transfer learning · AgileX and AIRBOT | progress score | Mixed |
Closest papers by meaning
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.88)
- Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning (ICML 2025 · similarity 0.88)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.88)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.88)
- SkiP: When to Skip and When to Refine for Efficient Robot Manipulation (NeurIPS 2026 · similarity 0.87)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · 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.
