Research map / Reinforcement learning / RL from feedback
PACT: Self-Evolving Physical Safety Alignment for Diffusion Policies in Embodied Manipulation
Lingxuan Wu, Zijian Zhu, Lizhong Wang, Chengyang Ying, Huayu Chen, Xiao Yang, Fangming Liu, Jun Zhu
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 |
|---|---|---|---|
| post-train with PACT | DP · RoboTwin 7 tasks avg | success rate: 45% -> 73% | Improved |
| post-train with PACT | DP · RoboTwin 7 tasks avg | safe rate: 41% -> 65% | Improved |
| post-train with PACT | DP3 · RoboTwin 7 tasks avg | success rate: 27% -> 43% | Improved |
| post-train with PACT | RDT-1B · RoboTwin 7 tasks avg | success rate: 56% -> 79% | Improved |
| post-train with PACT | pi0.5 · RoboTwin 7 tasks avg | safe rate: 37% -> 62% | Improved |
| use PACT distillation instead of PPO | DP · RoboTwin Pour Water | safe rate: 31% -> 64% | Improved |
| use PACT distillation instead of DIPO | DP · RoboTwin Pour Water | success rate: 67% -> 86% | Improved |
| use on-policy PACT instead of off-policy distillation on guided rollouts (ablation) | DP · RoboTwin Pour Water | success rate: 72% -> 86% | Improved |
| use curriculum distillation instead of direct distillation (ablation) | PACT · RoboTwin | training stability | Improved |
| increase Lagrange multiplier lambda too large (ablation) | PACT · RoboTwin | performance | Hurt |
| inject cost gradient over larger max diffusion time t_c (ablation) | PACT · RoboTwin | performance | Hurt |
| raise UTD ratio from 100 to 200 (288 rollouts) (ablation) | PACT · RoboTwin Handover Apple | success rate: 75% -> 45% | Hurt |
Closest papers by meaning
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.87)
- DexHandDiff: Interaction-aware Diffusion Planning for Adaptive Dexterous Manipulation (CVPR 2025 · similarity 0.87)
- Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.86)
- GENMANIP: LLM-driven Simulation for Generalizable Instruction-Following Manipulation (CVPR 2025 · similarity 0.86)
- ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY (ICLR 2025 · similarity 0.86)
- DeformGen: Dynamics-Based Topology Augmentation for Deformable Manipulation Policy Learning (NeurIPS 2026 · similarity 0.86)
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.
