Driving · 2026-11-10

Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents

Yifei Yao, Yude Zou, Xing Gao, Guanbo Jiao, Yukai Wang, Jianming Ma, Weinan Zhang, Chunhua Shen

arXiv:2611.11509PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

SimplEvo improves a frontier MLLM under direct control using two self-written, fixed-size text memories: a coordinate-free task memory from one demonstration and a bounded working memory refined after each training episode via single-turn trajectory reflection. No weight updates, post-trained policies, or privileged information are used, and both memories are fixed at test time. On EBench and RoboDojo, the authors report large success-rate gains for GPT-6-Astra.

Why it matters

The authors test whether performance gains from prior harness engineering (post-trained policies, RSI loops, etc.) exceed what the base MLLM can learn by retaining its own execution experience across episodes. Under a strict train–test protocol and an information protocol limited to RGB views and proprioception, SimplEvo shows that memory evolution alone can boost direct-control robot success and transfer the resulting working memory across MLLMs.

Method

  • Direct control remains unchanged; the MLLM issues all robot commands from RGB views and proprioception (no privileged depth/camera/object states).
  • Write a coordinate-free task memory once from a single demonstration: stage order plus a visual completion criterion (terminal-state anchoring).
  • After each training episode, rewrite a bounded working memory (≤32KB) in one reflection turn distilled from the executed trajectory and episode outcome; fix both memories for test time.

Limitation

Tasks without a demonstration, the eight Open tasks of RoboDojo, run without Mtask.

Abstract (from arXiv)

Recent studies of embodied agents based on frontier multimodal large language models (MLLMs) develop harness engineering around the MLLM, such as harnesses with post-trained policies and recursive self-improvement loops. Yet whether these additions exceed the native capability of the MLLM remains unexplored. We introduce SimplEvo to explore how far an MLLM can improve under direct control when it retains its execution experience in self-written text memories. It requires no weight updates, post-trained policies, or privileged information. The MLLM writes a coordinate-free task memory once from a single demonstration. Under a strict train-test protocol, it then creates a bounded working memory on training episodes only and refines it after each episode with lessons on procedure, completion, and recovery. Both memories stay fixed at test time. On EBench, SimplEvo lifts the success rate (SR) of GPT-6-Astra under direct control from 46.7% to 81.7%, and tops the RoboDojo leaderboard with 69.63% SR and a score of 75.32. Compared with memories that also store past visual frames, SimplEvo reaches a similar SR with text memory alone while using fewer tokens. With its memories fixed, SimplEvo also generalizes to unseen layouts and randomized scene configurations. Moreover, the working memory transfers across models: written by GPT-6-Astra, it can be reused as is by weaker MLLMs and improve their performance without any adaptation of their own.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First announced by arXiv: 2026-11-10 · Indexed by Pub-AI: 2026-11-10 UTC

preprint

Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents

Yifei Yao, Yude Zou, Xing Gao, Guanbo Jiao, Yukai Wang, Jianming Ma, Weinan Zhang, Chunhua Shen

First announced by arXiv: 2026-11-10 · Original record ↗

arxiv
2611.11509
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-10 · First detected: 2026-11-10 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.11509v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

Recent studies of embodied agents based on frontier multimodal large language models (MLLMs) develop harness engineering around the MLLM, such as harnesses with post-trained policies and recursive self-improvement loops. Yet whether these additions exceed the native capability of the MLLM remains unexplored. We introduce SimplEvo to explore how far an MLLM can improve under direct control when it retains its execution experience in self-written text memories. It requires no weight updates, post-trained policies, or privileged information. The MLLM writes a coordinate-free task memory once from a single demonstration. Under a strict train-test protocol, it then creates a bounded working memory on training episodes only and refines it after each episode with lessons on procedure, completion, and recovery. Both memories stay fixed at test time. On EBench, SimplEvo lifts the success rate (SR) of GPT-6-Astra under direct control from 46.7% to 81.7%, and tops the RoboDojo leaderboard with 69.63% SR and a score of 75.32. Compared with memories that also store past visual frames, SimplEvo reaches a similar SR with text memory alone while using fewer tokens. With its memories fixed, SimplEvo also generalizes to unseen layouts and randomized scene configurations. Moreover, the working memory transfers across models: written by GPT-6-Astra, it can be reused as is by weaker MLLMs and improve their performance without any adaptation of their own.