Research map / Reinforcement learning / Imitation learning
Learning Shared Safety Constraints from Multi-task Demonstrations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use 10 rather than 1 constraint-learning iteration | ICL · Position Constraint | |c* - c|: 0.052 -> 0.000 | Helped |
| Use 10 rather than 1 constraint-learning iteration | ICL · Position Constraint | J(πE, r) - J(π, r): 569.277 -> -72.149 | Helped |
| Use 10 rather than 1 constraint-learning iteration | ICL · Velocity Constraint | |c* - c|: 0.249 -> 0.006 | Helped |
| Use 10 rather than 1 constraint-learning iteration | ICL · Velocity Constraint | J(πE, c*) - J(π, c*): -244.25 -> -21.602 | Helped |
| Use multi-task demonstrations from all tasks to train one shared constraint network (ablation) | ICL · AntMaze, 10 tasks | constraint accuracy against maze walls | Helped |
| Use only single-task demonstrations to learn the constraint (ablation) | ICL · AntMaze | constraint generalization across tasks | Hurt |
| Aggregate constraints learned separately for each task by averaging or taking their maximum (ablation) | ICL · AntMaze, 10 tasks | constraint accuracy against maze walls | Hurt |
| Use Chou et al.'s constraint-learning approach rather than iterative ICL | Position Constraint · Position Constraint | J(πE, c*) - J(π, c*): 133.59 vs -1958.372 | Hurt |
| Use Chou et al.'s constraint-learning approach rather than iterative ICL | Velocity Constraint · Velocity Constraint | J(πE, c*) - J(π, c*): -21.602 vs -134.58 | Hurt |
| Add raw constraint value to the policy state (ablation) | CRL · continuous-control tasks | policy training success | Helped |
| Use a log-activation on top of the raw constraint value (ablation) | CRL and ICL · continuous-control tasks | training stability | Helped |
| Warm-start each inner-loop policy with behavior cloning from expert demonstrations (ablation) | ICL · continuous-control tasks | interaction required for inner optimization | Helped |
Closest papers by meaning
- Simplifying Constraint Inference with Inverse Reinforcement Learning (NeurIPS 2024 · similarity 0.91)
- When a Robot is More Capable than a Human: Learning from Constrained Demonstrators (ICLR 2026 · similarity 0.90)
- Understanding Constraint Inference in Safety-Critical Inverse Reinforcement Learning (ICLR 2025 · similarity 0.89)
- Handling Long and Richly Constrained Tasks through Constrained Hierarchical Reinforcement Learning (AAAI 2024 · similarity 0.87)
- Robust Inverse Constrained Reinforcement Learning under Model Misspecification (ICML 2024 · similarity 0.87)
- Safety Representations for Safer Policy Learning (ICLR 2025 · 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.