ReSYNC turns robot failures into reusable skills, accepted to CoRL 2026
DJiafei · x · 2026-10-13
Researchers from CMU, Princeton, MIT, UCSD, and Bosch AI introduced ReSYNC, accepted to CoRL 2026, which converts robot deployment failures into plannable skills.
- Recover: when Pick fails (block against a wall), GPU-parallel simulation spawns worlds from the failure state and PPO learns a relational Poke recovery policy
- Discover: counterfactual dreaming separates useful relations from accidental details, adding new concepts (e.g. Poke induces TopUp) to the planning model
- Plan: on new tasks, the robot reasons backward from the goal and schedules Poke first, avoiding failure entirely
- Validated across six physical domains; paper and code released
More from Embodied
- GeneralistAI expands GEN-1.5 access, betting on a brain for billions of robots — E0M · 2026-10-13
- Microduck Fight: a 25cm servo-duck fighting game where two neural policies duke it out — huggingface · 2026-10-13
- Foxglove founder: billions of autonomous machines are coming, and infra is the bottleneck — Sethwinterroth · 2026-10-13
- Ex-Boston Dynamics strategy chief's startup unveils Hugo, an interactive furry robot for theme parks — CyberRobooo · 2026-10-13
- 7 AI models flew a drone via phone MCP: Claude refused, Codex one-shotted it — Basic-Let6828 · 2026-10-13
- SEA-LLM from a Google internship brings egocentric spatial audio understanding to smart glasses — AcouIntel · 2026-10-13