Robotics · 2026-10-14
EmbodiedEvo: Self-improving Coding Agents for Physical Intelligence
Wenyi Li, Yizhe Chi, Youjie Zheng, Chunyu Liu, Xiaoqiu Wang, Mingju Gao, Zheng Jiang, Deyao Hong, Calvin Xiao, Kaisen Yang, Qinhuai Na
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
EmbodiedEvo is a self-improving coding agent for physical tasks that evolves an evolving harness of reusable “Skills” and “Lessons” from feedback on one Playground scene, then freezes it for Test agents on new layouts. On RoboDojo-Sim with Astra/xhigh, the authors report Average Score 44.48 and success rate 37.82%, surpassing all officially verified results. They also report cross-model transfer benefits and lower execution cost than using raw history.
Why it matters
The authors argue that physical intelligence benefits from “Code as World” and “Agent as Policy”: code makes control logic explicit, inspectable, and reusable, and an agent can revise code via feedback. EmbodiedEvo targets the issue that physical execution is failure-prone by storing reusable procedures (Skills) and applicability/failure guidance (Lessons) acquired from one scene and reused across new layouts and different executors.
Method
- Uses code as an explicit control representation with a two-level interface: environment primitives (L1) and composed Python procedures (L2) built from those primitives.
- Runs an agent-directed loop in a Playground persistent simulation: the agent edits submission/solution.py, executes, inspects feedback, and revises Skills and Lessons within the episode.
- After Playground, freezes the Skills/Lessons harness as read-only artifacts for Test agents on new layouts (no reset, no harness updates), enabling reuse across scenes and models.
Limitation
The authors state that fine contact control remains difficult.
Abstract (from arXiv)
Bringing general-purpose agents into the physical world calls for an executable representation of physical intelligence. Compared with vision-language-action (VLA) models, which map observations and language instructions directly to actions, code makes control logic explicit and reusable. Code can serve not only as a policy representation but also as the workspace in which the agent solves a task. Rather than following a prescribed workflow of planning, execution, and verification, the agent can autonomously inspect the state, revise and execute code, and use feedback to choose the next step. The success of this SWE-style loop in digital environments motivates its extension to physical tasks. Yet physical execution remains prone to recurring failures. To retain and refine experience for later scenes, we introduce EmbodiedEvo, where a coding agent evolves reusable Skills and Lessons through feedback on a single Playground scene, then freezes the resulting harness for diverse new scenes. On RoboDojo-Sim, EmbodiedEvo achieves an average score of 44.48 and a success rate of 37.82\%, surpassing all officially verified results. Paired tests show the largest gains on memory-dependent tasks and successful reuse of coordinated procedures, while fine contact control remains difficult. Experience from stronger models also improves weaker executors: Astra's experience raises Sol and Luna's success rates by 8.86 and 5.50 percentage points over no experience. Compared with raw interaction history, Skills and Lessons achieve comparable task performance at 53.0\% lower execution cost.
Related papers
- AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents
- RoboBridge: A Self-Evolving Embodied Agent Framework for Sim-to-Real Transfer
- SEAR: Self-Evolving Agentic Robotics
- Explore, Execute, Evolve: A Skill Acquisition and Reuse Loop for Embodied Agents
- EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-13 · Indexed by Pub-AI: 2026-10-14 UTC
EmbodiedEvo: Self-Evolving Coding Agent For Physical Intelligence
Preprint posted: 2026-10-13 · Original record ↗
- arxiv
- 2610.18867
Explicit source versions (1)
- v2 · 2026-10-14
- Preprint available
Source date: 2026-10-13 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
- Explicit source version detected · v2
Source date: 2026-10-14 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
Abstract · arxiv
Coding agents can construct executable programs and revise them through feedback, offering a foundation for embodied control without target-task fine-tuning. Through a programmatic robot interface, the agent observes the scene, executes control programs, and adapts them to the resulting physical state, serving as the closed-loop policy. However, this within-episode adaptation does not automatically preserve experience across scenes. Our trajectory analysis shows that agents spend interaction and inference budgets establishing effective operations and identifying their preconditions, while fresh agents cannot reuse these discoveries. This motivates EmbodiedEvo, a self-evolving coding agent that distills physical feedback into reusable code Skills and diagnostic Lessons. Experience is acquired on one Playground scene per task, then frozen for independent agents in unseen scenes, without updating model weights. On RoboDojo-Sim, EmbodiedEvo achieves an average score of 44.48 and a success rate of 37.82\%, surpassing all officially verified results. Paired tests show the largest gains on memory-dependent tasks and successful reuse of coordinated procedures, while fine contact control remains difficult. Experience from stronger models also improves weaker executors: Astra's experience raises Sol and Luna's success rates by 8.86 and 5.50 percentage points over no experience. Compared with raw interaction history, Skills and Lessons achieve comparable task performance at 53.0\% lower execution cost.