Recursia's inner-outer loop pushes embodied RSI, crushing π₀ on real-robot compositional tasks

机器之心 · wechat · 2026-10-14

A team from Beijing Jiaotong University, Peking University, NYU, and USC proposes Recursia, framing compositional generalization within embodied recursive self-improvement: an inner loop handles observation, decisions, and code execution, while an outer loop proposes, verifies, and retains edits to a shared executable stack so improvements persist across tasks.

Key results

Architecture

The article also surveys RSI efforts at Anthropic (Claude writes over 80% of internal merged code), OpenAI (RSI Index in GPT-5.6 notes), and NVIDIA's ASPIRE/ENPIRE, and notes that multimodal-LLM-driven arms still bottleneck on 3D geometry and contact-state perception in fine manipulation.

Original post →

More from Embodied

Embodied channel →