BOSS: LLM-guided agents grow skill libraries to zero-shot solve long-horizon tasks
chris_j_paxton · x · 2026-08-15
Jesse Zhang (USC) resurfaces memories of waiting days to hit 500k timesteps on ALFRED/Ai2Thor, linking to the paper BOSS (BootStrapping your Own SkillS) — a CoRL 2023 Oral (top 6.6%) from USC, Google, National Taiwan University and KAIST.
The method has two phases:
- Pre-training a language-conditioned skill policy: a policy and critic are trained with offline RL (IQL in the experiments, chosen for performance and online fine-tuning friendliness) on easy-to-collect, short-horizon, language-annotated trajectories.
- Skill bootstrapping: the agent practices in the target environment, tries new chains of skills, and adds successful long-horizon behaviors back into its repertoire for further bootstrapping.
Once the skill library is ready, conditioning the policy on language descriptions allows zero-shot execution of novel, complex long-horizon tasks.
More from Embodied
- TAU Robotics cleaning service shows teleoperated future — Distinct-Question-16 · 2026-08-15
- Low loss doesn't equal high success rate: robotics needs better metrics — mathildepapillo · 2026-08-15
- MagicMirror² turns mirrors into smart assistants with modules — tom_doerr · 2026-08-15
- Developer shares prototype footage of small humanoid robot 'The Android' — Jasonio · 2026-08-15
- Training General Robots for Any Task with Physical Intelligence — davidyin44 · 2026-08-15
- PIRT: Run a Linux Agent Directly on Your Android Phone — Glad_Raspberry_6795 · 2026-08-15