rsi.watch

Xiaomi Embodied Intelligence Team (31 named authors) · 2026-10-26

OmniBrain-0: A Unified Agentic Model for Embodied Intelligence

“RSI further raises success to 69.3% at the same average completion time, improving robustness without additional execution time.”
Who
Xiaomi Embodied Intelligence Team (31 named authors)
Date
2026-10-26
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ risk of bias. The usable demonstrations, RSI included, are counted only from LIBERO/LIBERO-Pro, RoboTwin 2.0, RoboCasa365 and R2R-CE, benchmarks in the results tables. The paper does not say that the evaluated tasks or seeds were kept out of RSI collection; RoboCasa365 is evaluated on held-out seeds. “The resulting corpus combines successful trajectories from the initial expert collection with additional verified successful trajectories collected through the RSI Data Flywheel.”
  • ↓ imprecise. Each RSI gain in the tables, such as 83.8% to 84.5% on LIBERO-Plus over 840 episodes and 65.5% to 69.3% on LIBERO-Pro, is reported as a single number per model, with no confidence intervals or repeated training runs. R2R-CE uses 100 episodes, where SR goes from 65.0% to 66.0%. “We evaluate a fixed subset of 100 episodes from the Val-Unseen split.”
how we rate evidence
Replication
No rerun found. The repository linked from the project page says "Code: release soon." We found no rerun by another group. Searched up to 2026-10-29. all claims
Checked
2026-10-29, against the source page
Links
Source Paper Code

Also on the page

  • OmniBrain-0 is a Qwen3.5-9B planner that calls manipulation and navigation tools. It is trained by embodied pretraining, distillation of teacher trajectories (GPT-5.6-Sol and GLM-5.3-Flash) into domain experts, and multi-teacher on-policy distillation (MOPD) of those experts into one model. RSI is a further stage: a Meta-Agent (GPT-6.1-Sol) revises rubrics and step-level reminders from execution evidence, verified trajectories are added to the expert corpora, and the retrained experts are distilled again into OmniBrain-0, which collects the next round.
  • The quote is the RSI row of the LIBERO-Pro table, against 65.5% for the OmniBrain-0 (9B) row; the paper describes the RSI row as "the model iterated through the RSI Data Flywheel". Overall scores, 9B row to RSI row: 83.8% to 84.5% on LIBERO-Plus, from 39.6% to 43.2% on RoboCasa365, from 88.2 to 89.4 on RoboTwin 2.0, and from 65.0% to 66.0% SR on a 100-episode R2R-CE Val-Unseen subset. RSI contributes 1,142 demonstrations on top of an initial agentic corpus of 3,846.
  • Most of the gain over the few-shot Qwen3.5-9B backbone comes from the training stages before RSI: on RoboCasa365 the backbone scores 12.2% and OmniBrain-0 (9B) 39.6%; on R2R-CE SR goes from 34.0% to 65.0%. The backbone row uses the Experience Library; the OmniBrain-0 rows are evaluated "without an external Experience Library".
  • The paper gives no number of RSI rounds. Its reminder ablation, reminders against none with the checkpoint fixed, is reported only in a figure. The expert demonstrations come from LIBERO/LIBERO-Pro, RoboTwin 2.0, RoboCasa365 and R2R-CE (2,349, 513, 880 and 1,246), 3,846 plus 1,142 in all, so the RSI trajectories come from benchmarks the model is evaluated on; the paper does not say which tasks or seeds RSI collected on.
  • The authors' limits: "OmniBrain-0 remains dependent on the coverage and reliability of its low-level skills", "gaps persist on unseen task compositions and spatial localization", and "Real-world evaluation also covers a limited set of tasks."

In the self-improving AI catalog

  • OmniBrain-0: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that OmniBrain-0 cites, oldest first. Under each: where that work sits on the map, then where OmniBrain-0 sits (changes in bold), and what OmniBrain-0's own paper says about it. Its whole family on the citation map

Cites Reflexion 2023

← Ledger Report an error