Research map / Reinforcement learning / Visual and representation RL
Autonomous Play with Correspondence-Driven Trajectory Warping
William Liang, Sam Wang, Hung-Ju Wang, Osbert Bastani, Yecheng Jason Ma, Dinesh Jayaraman
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 |
|---|---|---|---|
| replace keypoint-correspondence imitation policy (KALM) with trajectory warping | Tether · Real-world Pineapple to Bowl | success rate: 0.0 -> 1.0 | Improved |
| increase source demos from 1 to 10 (ablation) | Tether · Real-world Wiping with Cloth | success rate: 0.1 -> 0.6 | Improved |
| increase source demos from 1 to 10 (ablation) | Tether · Real-world Pineapple to Cup | success rate: 0.5 -> 0.9 | Improved |
| increase source demos from 1 to 10 (ablation) | Tether · Real-world Inserting Coffee | success rate: 0.2 -> 0.4 | Improved |
| increase source demos from 5 to 10 (ablation) | Tether · Real-world Opening Cabinet | success rate: 0.3 -> 0.6 | Improved |
| increase source demos from 1 to 10 on easy pick-and-place (ablation) | Tether · Real-world Strawberry to Bowl | success rate: 0.9 -> 0.9 | No clear effect |
| use Tether trajectory warping instead of zero-shot pi0 | Tether vs pi0 · Real-world Bowl to Shelf | success rate: 0.0 -> 0.9 | Improved |
| finetune pi0 on play-generated demos instead of 10 human demos | pi0-FAST-DROID · Real-world Pineapple to Bowl | success rate: 0.0 -> 0.8 | Improved |
| increase play-generated finetuning data from 50 to all demos (ablation) | pi0-FAST-DROID · Real-world Bowl to Shelf | success rate: 0.1 -> 0.4 | Improved |
| use Gemini Robotics-ER 1.5 alone instead of GPT-4.1 with correspondence-based verification for success evaluation (ablation) | Tether play success evaluator · Autonomous play annotated attempts | precision: 97.7 -> 98.4 | Improved |
| use Gemini Robotics-ER 1.5 instead of Gemini-2.5-Flash for task planning (ablation) | Tether play planner · Autonomous play annotated attempts | task planning accuracy: 94.1 -> 95.2 | Improved |
| use Tether policy instead of Diffusion Policy trained on 141-202 human demos for play states | Autonomous play policy · Play-induced initial states (20 per task) | success rate | Improved |
Closest papers by meaning
- OSVI-WM: One-Shot Visual Imitation for Unseen Tasks using World-Model-Guided Trajectory Generation (NeurIPS 2025 · similarity 0.87)
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (NeurIPS 2026 · similarity 0.87)
- RoboDreamer: Learning Compositional World Models for Robot Imagination (ICML 2024 · similarity 0.87)
- AffordDP: Generalizable Diffusion Policy with Transferable Affordance (CVPR 2025 · similarity 0.86)
- RoboScape: Physics-informed Embodied World Model (NeurIPS 2025 · similarity 0.86)
- DynScene: Scalable Generation of Dynamic Robotic Manipulation Scenes for Embodied AI (CVPR 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.
