SenseNova-RoboRSI-27B-Preview

简体中文 · GitHub · Demos · Quickstart

SenseNova-RoboRSI-27B-Preview is an open-weight 27B vision-language model for embodied agents. Through joint model–Harness RSI, it learns from validated interaction experience, developing capabilities for planning, action decisions, and feedback-based correction in simulation and on real robots.

This guide covers model deployment and benchmark reproduction on LIBERO-Pro and RoboCasa365 using RoboRSI Harness. RoboDojo reproduction instructions will be released alongside upcoming RoboDojo support in RoboRSI. A detailed technical report is coming soon.

• RoboDojo: Achieves an average score of 26.13 and a task success rate of 20.83% with VLA-assisted execution. Its average score is just 2.84 points below GPT-6 Astra's score in the RoboDojo team's evaluation. The model actively inspects visual details, reasons about spatial relationships, and corrects actions based on execution feedback during complex manipulation tasks.
• LIBERO-Pro: Reaches 92.5% success with VLA-assisted execution (System 2 + System 1) and 95.0% with direct VLM control (System 2), ranking second to GPT-6 Astra among the compared models in both settings. These results demonstrate embodied planning and decision-making across different execution architectures.
• RoboCasa365: RSI training raises success from 26.0% to 46.5% over the Qwen3.8-27B base model, while reducing average output tokens from 59.5K to 15.9K, a 73.3% reduction. Learning from interaction experience improves both task performance and inference efficiency.

Demonstrations

Real-robot demonstration

Simulation · RoboDojo

Simulation · LIBERO

Simulation · RoboCasa365

Quickstart

1. Deploy the model

Install vLLM

In your inference environment, install the CUDA 12.9 build with a compatible NVIDIA driver:

pip install "vllm==0.21.0" huggingface_hub \
  --extra-index-url https://download.pytorch.org/whl/cu129

See the vLLM installation guide for other CUDA builds.

Download the model

Download the full checkpoint, including the tokenizer, processor, generation configuration, and chat template:

export MODEL_DIR="$PWD/models/sensenova-roborsi-27b-preview"
hf download sensenova/sensenova-roborsi-27b-preview \
  --local-dir "$MODEL_DIR"

For an existing local checkpoint, set MODEL_DIR to its absolute path.

Start the service

export CUDA_VISIBLE_DEVICES=0,1,2,3

vllm serve "$MODEL_DIR" \
  --served-model-name sensenova-roborsi-27b-preview \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype bfloat16 \
  --tensor-parallel-size 4 \
  --max-model-len 262144 \
  --generation-config auto \
  --chat-template "$MODEL_DIR/chat_template.jinja" \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --structured-outputs-config '{"backend":"xgrammar","disable_any_whitespace":true}' \
  --enable-prefix-caching \
  --mamba-cache-mode align

We recommend medium reasoning effort for this preview. Set chat_template_kwargs.reasoning_effort explicitly in vLLM Chat Completions requests, as shown below. The bundled chat template defaults to xhigh when this setting is omitted.

Check the endpoint

From another terminal on the same machine, check service health and send a short robot-planning request:

curl --fail http://127.0.0.1:8000/health
curl --fail http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "sensenova-roborsi-27b-preview",
    "messages": [{
      "role": "user",
      "content": "A robot needs to place a mug on a coaster. Give a short plan and explain how to verify that the mug is stable before releasing it."
    }],
    "max_tokens": 8192,
    "chat_template_kwargs": {"reasoning_effort": "medium"},
    "stream": false
  }'

2. Connect RoboRSI Harness and run evaluations

Install RoboRSI Harness and the evaluation environment following the model evaluation guide, then use the commands below:

export OPENAI_BASE_URL='http://127.0.0.1:8000/v1'
export OPENAI_API_KEY='EMPTY'
export MODEL_NAME='sensenova-roborsi-27b-preview'

# LIBERO-Pro: VLM + VLA (System 2 + System 1)
# Set up the tool environment, download weights, and deploy VLA/SAM3 in the background; no automatic evaluation
python scripts/deploy/libero_tools.py serve \
  --model-config configs/models/roborsi_27b.yaml \
  --work-dir work/model_eval/libero_tools &

# Wait for Services ready before running the evaluation command below (default URLs shown)
ROBORSI_VLA_URL=http://host.docker.internal:24001 \
ROBORSI_SAM3_URL=http://host.docker.internal:24002 \
python scripts/evaluation/evaluate_model.py --benchmark libero_pro \
  --benchmark-config configs/benchmarks/libero_pro_80.yaml \
  --model-config configs/models/roborsi_27b.yaml \
  --agent-config configs/agents/libero/sensenova_roborsi_27b_vla_sam3.yaml

# LIBERO-Pro: Direct VLM control (System 2); skip the VLA/SAM3 deployment command above
python scripts/evaluation/evaluate_model.py --benchmark libero_pro \
  --benchmark-config configs/benchmarks/libero_pro_80.yaml \
  --model-config configs/models/roborsi_27b.yaml \
  --agent-config configs/agents/libero/sensenova_roborsi_27b_no_tools.yaml

# RoboCasa365
python scripts/evaluation/evaluate_model.py --benchmark robocasa365

# RoboDojo: Evaluation support will be available in a future RoboRSI Harness update

Built on the RoboRSI Harness

SenseNova-RoboRSI-27B-Preview runs within the RoboRSI Harness, the shared foundation for task execution, evaluation, and iterative improvement. This connects the model's learned capabilities to robot actions and turns new interactions into evidence for the next RSI round.

• Composable execution. Connect models, prompts, skills, and tools/MCP; coordinate VLA skills, scripted actions, or end-effector control through a common agent loop.
• Reproducible evaluation. Benchmark adapters preserve native execution and scoring, while configurable context, caching, and batch execution support repeatable experiments across environments.
• Experience for continued learning. Full traces and exported training episodes connect execution to model training; versioned retesting identifies which Harness changes to carry forward.

We develop and maintain SenseNova-RoboRSI on GitHub as the shared foundation for community deployment, reproduction, and continued training. Follow the repository for the Harness, evaluation data, intermediate results, and release updates.

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, Quanzhou Li, Ziyi Shang, Wanqi Yin, Xiaowan Li, Donghao Wu, Tianbo Zhong
• Advisor: Dahua Lin, Xiaogang Wang, Ziwei Liu, Lewei Lu, Lei Yang

* denotes team leads.

License

The model weights and model card are released under the Apache License 2.0. See LICENSE for the full license and NOTICE for upstream attribution and a summary of our modifications.

The model is derived from Qwen3.8-27B, whose weights were further post-trained by SenseNova for embodied reasoning and interaction. The upstream license and copyright notice are retained in LICENSE.

External models, simulators, and datasets retain their respective licenses.