Skip to content

About

A unified framework for building, evaluating, and recursively improving embodied agents. 🚀 Code release coming soon.

Resources

Contributing

Stars

830 stars

Watchers

29 watching

Forks

Repository files navigation

SenseNova-RoboRSI: Recursive Self-Improvement for Embodied Agents

Build, evaluate, and improve embodied agents with a shared runtime.

Python 3.10–3.12 License: Apache 2.0 Status: Pre-Alpha

Report: Has GPT-6 Achieved Embodied Intelligence? Model: SenseNova-RoboRSI-27B-Preview

English · 简体中文 · Quick Start · Report · RoboRSI-27B-Preview

News

  • 2026-11-16 — SenseNova-RoboRSI is now listed on the official RoboDojo Agent leaderboard, with an average score of 57.10 and a success rate of 49.17% under Avg. (w/o Open).

  • 2026-11-15 — Added ICL package construction and verification tools, plus demonstration-learning and frozen-package evaluation guides for BEHAVIOR-1K, Bench2Dex and DexJoCo. ICL guide · Source coverage.

  • 2026-11-03 — Initial code and usage guides for LHPR-VLN and VLN-CE (R2R / RxR) are available.

  • 2026-10-30 — Initial code and usage guides for RoboLab are available.

  • 2026-10-27 — Our LIBERO-PRO result (98.25% success rate) is now officially ranked No. 1 on the Papers With Code leaderboard.

Earlier updates
  • 2026-10-26 — SenseNova-RoboRSI is now ranked No. 1 on the RoboLab leaderboard.

  • 2026-10-22 — SenseNova-RoboRSI-27B-Preview now runs across different environments, including on real robots and in simulation! 🤗 Model weights and usage guides are now available. The technical report is coming soon.

  • 2026-10-22 — Initial code and usage guides for RoboCasa365 Target50 are available.

  • 2026-10-21 — Added benchmark images and ICL experience packages to BenchZoo, including image references and package download links.

  • 2026-10-20 — Initial code and usage guides for BEHAVIOR-1K, Bench2Dex and DexJoCo are available.

  • 2026-10-16 — Initial code and usage guides for LIBERO, LIBERO-PRO, LIBERO-Long++ and LIBERO-Ultra-Long are available. See the getting started guide.

  • 2026-10-16 — SenseNova-RoboRSI has been used in the empirical report Has GPT-6 Achieved Embodied Intelligence?.

  • 2026-10-09 — SenseNova-RoboRSI achieves an average score of 60.08 and a success rate of 54.17% on RoboDojo, a gain of +31.11 points over the GPT-6 Astra baseline (28.97) in our original evaluation. See Harness Results for the original comparison and the subsequently published official Agent-track results.

Overview

Connect your model, skills, and robot tools through a shared runtime, customize prompts and context, and evaluate policies in environments that retain their native execution and scoring logic.

On top of this runtime, recursive self-improvement (RSI) uses execution traces and task feedback to improve the Harness while keeping model parameters fixed. Candidate changes are retested, and validated improvements are carried into the next agent version, connecting task execution with continuous improvement.

🎉 SenseNova-RoboRSI-27B-Preview extends this work to model training. Validated interaction experience becomes multimodal supervision for goal understanding, action decisions, and feedback-based correction. Updated models return to the same runtime for evaluation and further exploration, connecting model training and Harness improvement through shared experience and joint retesting.

SenseNova-RoboRSI agent interfaces, execution loop, benchmark integration, and RSI workflow

Core capabilities

Harness: A shared runtime

  • Flexible agent interfaces. Connect models, prompts, skills, and tools/MCP; configure context and caching to suit each task.
  • Unified execution and evaluation. Run VLA skills, scripted actions, or end-effector tools across benchmarks with native feedback and scoring.
  • Experiment records and training data. Record full traces, export training episodes, and run batch evaluations with concurrency and recovery.

RSI: Continuous Harness improvement

  • Learn from execution feedback. Refine action prediction, planning, feedback, and memory organization while keeping model parameters fixed.
  • Retest and retain improvements. Track agent versions, evaluate candidate changes under the same task conditions, and carry validated improvements forward.

Quick Start

Use Linux/WSL, Python 3.10–3.12 and Docker; prepare GPU support where required. From the repository root:

python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[runtime]'

Installation and environment prerequisites: getting started.

GPT-6 Astra: evaluate all released benchmarks

Prepare the environments and Agents once using the suite guide, and set their paths/services in the suite configuration. Then start all released benchmark cohorts with:

export OPENAI_API_KEY='YOUR_API_KEY'
python scripts/evaluation/evaluate_gpt6.py

The script runs each benchmark's existing Agent and reports its native scores separately. See the guide for coverage, frozen ICL, offline plans and results.

SenseNova-RoboRSI-27B-Preview: use the existing evaluation entry

Deploy the model, then use the existing deployed-model evaluator:

export OPENAI_BASE_URL='http://127.0.0.1:8000/v1'
export OPENAI_API_KEY='EMPTY'
export MODEL_NAME='sensenova-roborsi-27b-preview'
python scripts/evaluation/evaluate_model.py --benchmark robocasa365

LIBERO-Pro direct control and VLA/SAM3 commands remain in the model guide. Other models and advanced options: model evaluation.

BenchZoo

Available benchmark integrations and upcoming releases. Supported services share configurable EmbodiedAgent presets; select your model and follow the corresponding benchmark directory for setup and evaluation.

ICL experience packages are hosted in SenseNova-RoboRSI-Packages. The links below open each released benchmark's versioned package directory; follow the package documentation for download, verification and usage.

Prebuilt images are hosted on GHCR. Use docker pull <image> with the corresponding image reference from the table.

Benchmark Status ICL experience package Docker image
RoboDojo Coming soon — —
RoboCasa365 Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-robocasa365@sha256:28d210a42597f88c7af23b63f88fefc587edc901e268a0339cbfbf6ea1814cb5
LIBERO-Pro Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-libero-pro@sha256:a83417f1459a4c555c457caa82ef00a5de50191e3d8deb943894fa9285cff172
LIBERO-Long++ Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-libero-long-plus-plus@sha256:814139261e914b114cb89a359c6a39d4f409a556f79928e986b6ade3dfe7dc7a
LIBERO-Ultra-Long Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-libero-ultra-long@sha256:814139261e914b114cb89a359c6a39d4f409a556f79928e986b6ade3dfe7dc7a
RoboLab Released — ghcr.io/opensensenova/sensenova-roborsi-robolab@sha256:519585576bb09c5fc4a09e6255e30d74189b86b0f8c526d3fe8d99c5031da04f
Bench2Dex Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-bench2dex@sha256:edb0701c6ea3904c53a64820775336b998e6a3d0a9ac8b3b5be2caf824db3a1b
DexJoCo Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-dexjoco@sha256:19fa30540e3956ae460efa0a4b473586637cb330d863ea65ba0fd6318c376d4d
LHPR-VLN Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-lhvln@sha256:7d72dd910d5c9048ca6ec40ad158214497c3a9ac276641df1429314f2d7a3695
VLN-CE (R2R / RxR) Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-vln-ce@sha256:f54607562ca465101d6e1091f97d776755c2b00517f1852643bd8e0161c0e787
SIMPLE Coming soon — —
BEHAVIOR-1K Released v1.0.0 ghcr.io/opensensenova/sensenova-roborsi-behavior-1k:v1.0.0-cu128

Harness Results

Each table shows SenseNova-RoboRSI alongside selected reference results. Higher is better.

Official Agent leaderboard: SenseNova-RoboRSI is listed with an average score of 57.10 and a success rate of 49.17% under Avg. (w/o Open), the equal-weight mean of Generalization, Precision, Long-Horizon and Memory (accessed 2026-11-16).

The official entry carries a * annotation for deviations from the latest Agent Track Rules. The organizers note that the starred entries were evaluated before those rules were released.

Original evaluation: On RoboDojo-Sim, SenseNova-RoboRSI achieves an average score of 60.08 and a success rate of 54.17%. Compared with the GPT-6-Astra baseline, this is a gain of 31.11 score points and 31.69 percentage points in success rate. The table below preserves that original evaluation and its reference results.

Method Average score ↑ Success rate ↑
SenseNova-RoboRSI / GPT-6 Astra (ours) 60.08 54.17%
EmbodiedEvo 45.08 38.95%
VPP2 39.26 32.26%
PhysicalRSI 36.27 31.38%
Awomo-0.5 35.34 29.64%
Simate-beta 33.95 27.96%
GPT-6-Astra (baseline) 28.97 22.48%

Reference results for the original evaluation: RoboDojo simulation leaderboard (updated 2026-10-14, accessed 2026-10-15).

Method Success rate ↑
SenseNova-RoboRSI / GPT-6 Astra (ours) 94.50%
RPent / GPT-6 Astra 92.63%
RPent / Opus-4.7 82.40%
RPent / GPT-5.5 75.13%
RPent Flash Mode / Molmo2-8B 72.63%

The reference scores above, including GPT-6 Astra's 92.63%, are reported by RPent on Task/Swap suites. The official LIBERO-PRO leaderboard reports a different aggregate across five perturbation types.

Method Evaluation setting Success rate ↑
SenseNova-RoboRSI / GPT-6 Astra (ours) Target kitchens + task memory 64.60%
Paimon-0 Pretraining kitchens 58.10%
Xiaomi-Robotics-1 Pretraining kitchens 57.40%
Phasor-m7 Pretraining kitchens 56.80%
ProWAM Pretraining kitchens 48.10%

Reference results: official RoboCasa leaderboard. SenseNova-RoboRSI uses the Target50 setting; the official entries use pretraining kitchens. These results provide context across different evaluation settings and do not establish a ranking under the same protocol.

SenseNova-RoboRSI-27B-Preview

SenseNova-RoboRSI-27B-Preview is an open-weight 27B embodied VLM that extends RoboRSI from Harness improvement to model training. Through joint model–Harness RSI, it learns from validated interaction experience, bringing goal understanding, action decisions, and feedback-based correction into the model itself.

The model supports VLA-assisted execution and direct VLM control, with evaluations on RoboDojo, LIBERO-Pro, and RoboCasa365, and demonstrations in simulation and on real robots. The open weights provide a starting point for community deployment, reproduction, and continued training.

SenseNova-RoboRSI-27B-Preview results on RoboDojo, LIBERO-Pro, and RoboCasa365

Model weights are available on 🤗 Hugging Face. See the model guide for detailed results, demos, and usage instructions.

Release Plan & Roadmap

Release plan

Research roadmap

  • Experience retrieval. We plan to organize successful strategies, failure cases, and execution traces into a searchable experience library.
  • Transfer across tasks and environments. We will investigate whether tool combinations, planning strategies, and feedback mechanisms validated in one setting can provide useful starting points in another.
  • More robot platforms and real-world scenarios. We aim to extend integration and evaluation to more robot embodiments and physical settings, and examine the reliability of feedback-driven improvement in those environments.
  • Joint model–Harness RSI. Continue converting validated interaction experience into training supervision, retest model and system changes together, and measure capability and efficiency across improvement rounds.

Community & Acknowledgements

We welcome collaboration with teams working on foundation models, VLA policies, simulation platforms, and real robots.

  • Contributions: We welcome new integrations, evaluation tasks, reproducible failure cases, and suggestions for Harness design. See the contribution guide for the current release scope and checks.
  • Questions & collaboration: Open an issue describing your platform, use case, and proposed contribution.
  • Release updates: Watch the SenseNova-RoboRSI repository.
  • Acknowledgements: We thank RoboDojo, LIBERO-PRO, RoboCasa, RPent, and the broader open-source community supporting embodied-agent research.
  • Licenses: External models, simulators, and datasets retain their respective licenses.

This achievement reflects the joint team's continued exploration and close collaboration in embodied AI and agent systems. SenseNova-RoboRSI has been used in the published report Has GPT-6 Achieved Embodied Intelligence?. We will continue to advance and validate these technologies.

Contributors

  • Project leads: Hanming Deng, Zhongang Cai, Haiwen Diao, Zhi Hou, Zhonghua Wu, Qingyi Tao
  • Model team: Silei Wu*, Weiye Xu, Bo Liu, Pengfei Liu, Tailai Chen, Chenjing Ding, Jiahao Wang, Changxu Wei, Zongpu Zhang
  • Harness team: Jiapeng Luo*, Yuyao Shao, Linjun Dai, Ziyi Shang, Wanqi Yin, Xiaowan Li, Donghao Wu, Tianbo Zhong
  • Collaborators: Quanzhou Li, Han Chong Chua, Ruisi Wang, Xiaoming Wu, Tianqi Liu, Jing Zuo, Hongcan Xiao, Qinyue Li
  • Advisor: Dahua Lin, Xiaogang Wang, Ziwei Liu, Lewei Lu, Lei Yang

* denotes team leads.

Citation

If you use SenseNova-RoboRSI in your research, please cite the following paper:

@misc{sensenova2026gpt6astraachievedembodiedintelligence,
      title={Has GPT-6-Astra Achieved Embodied Intelligence?},
      author={SenseNova and Joint Research Team and Zhongang Cai and Ziang Cao and Tailai Chen and Han Chong Chua and Linjun Dai and Hanming Deng and Hokin Deng and Haiwen Diao and Chenjing Ding and Weichen Fan and Xiao Fu and Zerui Gong and Chenyang Gu and Zexin He and Zhi Hou and Bo Li and Qinyue Li and Quanzhou Li and Xiaowan Li and Zhichao Liao and Dahua Lin and Zilong Lin and Bo Liu and Chunxiao Liu and Pengfei Liu and Tianqi Liu and Ziwei Liu and Chen Change Loy and Lewei Lu and Jiapeng Luo and Yuexiao Ma and Liang Pan and Xingang Pan and Oscar Qian and Ziyi Shang and Yuyao Shao and Dawei Su and Dacheng Tao and Qingyi Tao and Quan Wang and Ruisi Wang and Xiaogang Wang and Ziyi Wang and Beichen Wen and Donghao Wu and Jiaxu Wu and Silei Wu and Xiaoming Wu and Zhonghua Wu and Hongcan Xiao and Weiye Xu and Atsushi Yamashita and Lei Yang and Zhe Yang and Baiqiao Yin and Wanqi Yin and Hanfei Zhao and Tianbo Zhong and Long Zhuo and Jing Zuo},
      year={2026},
      eprint={2610.22471},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.22471},
}

About

A unified framework for building, evaluating, and recursively improving embodied agents. 🚀 Code release coming soon.

Resources

Contributing

Stars

830 stars

Watchers

29 watching

Forks

Releases

Packages

Contributors

Languages