SSyncDev

SenseNova-RoboRSI-27B-Preview

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

SenseNova-RoboRSI-27B-Preview is SenseTime's open 27B vision-language model for embodied agents, post-trained from Qwen3.8-27B by recursive self-improvement.

Key specifications

Specification score
34.8
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights under Apache 2.0; no hosted price.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input

Produces

  • Text output

Capabilities

ReasoningFunction callingVisionOpen weights

What is SenseNova-RoboRSI-27B-Preview?

SenseTime's RoboRSI project bets that a robot model should get better by learning from its own validated attempts, and that the harness running it should evolve alongside. SenseNova-RoboRSI-27B-Preview is the first public checkpoint of that loop: a 27B vision-language model, post-trained from Qwen3.8-27B, that plans, picks actions and corrects itself from execution feedback in simulation and on real robots.

It is a preview. The technical report has not been published yet.

Results SenseTime highlights

Benchmark Setting Reported result
LIBERO-Pro VLA-assisted (System 2 + System 1) 92.5% success
LIBERO-Pro direct VLM control (System 2) 95.0% success
RoboCasa365 vs Qwen3.8-27B base 26.0% to 46.5% success
RoboDojo VLA-assisted 26.13 average score, 20.83% task success

The RoboCasa365 row carries a second, quieter win. Average output fell from 59.5K tokens to 15.9K, a 73.3% cut, so the trained model reasons less and succeeds more. On LIBERO-Pro and RoboDojo, SenseTime places it second to GPT-6 Astra in its comparisons, 2.84 points behind on RoboDojo. All figures are SenseTime's.

The harness is part of the product

The model is designed to run inside RoboRSI Harness, which SenseTime maintains on GitHub. It connects models, prompts, skills and tools (including MCP), and can drive VLA skills, scripted actions or end-effector control through one agent loop. It keeps native scoring for benchmark adapters, and exports full traces as training episodes, which is how the self-improvement rounds are fed. Reproduction guides currently cover LIBERO-Pro and RoboCasa365; RoboDojo support is still to come.

Deployment notes

  • Serve with vLLM (SenseTime documents the CUDA 12.9 build).
  • Download the full checkpoint, including processor, generation config and chat template.
  • Set chat_template_kwargs.reasoning_effort to medium, which SenseTime recommends. If you omit it, the bundled template defaults to xhigh, which will cost far more tokens.
  • Weights are Apache 2.0, with the upstream Qwen licence notice retained; simulators and datasets keep their own licences.

Where it fits

This is a high-level planner, the "System 2" in a two-tier setup. If you already run a VLA policy for motor control, it can sit above it; if not, the direct-control LIBERO-Pro result suggests it can act alone in simple scenes. Before committing, run your own task set through the harness at medium and at low effort and compare success rate against tokens per episode, since effort level drives cost directly.

Related embodied models include Tencent's Hy-Embodied-VLM-1.0 and Google's Gemini Robotics-ER 1.6 Preview. SenseTime's other open releases include SenseNova-Loop-9B-Preview. More models are under reinforcement learning.

How to evaluate SenseNova-RoboRSI-27B-Preview

Compare the model on the work you need to complete.

Tasks and inputs

SenseNova-RoboRSI-27B-Preview is listed for Image-text-to-text, Reinforcement learning. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Qwen3.8-27B, post-trained by SenseTime for embodied reasoning and interaction.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

runway

Praxis-1

Praxis-1 is Runway's open 27B world action model for robotics, predicting video and robot actions together; Apache 2.0, LeRobot and vLLM-Omni support.

— context

NVIDIA

Metatron-v1

Metatron-v1 is NVIDIA's open unified humanoid policy for the Unitree G1 that walks, interacts with scenes and recovers from falls, even in a browser.

— context

OpenBMB

MiniCPM-RobotManip

MiniCPM-RobotManip is OpenBMB's open 1.5B vision-language-action model for robot manipulation, keeping a minute of visual memory on-device.

— context

OpenBMB

MiniCPM-RobotTrack

MiniCPM-RobotTrack is OpenBMB's compact vision-language-action policy on MiniCPM4-0.5B that follows a described person on a Unitree Go2 at 5+ FPS on board.

— context

DocPereira

DocPereira/PEAL_V4_LHP_Zero_Entropy_Controlled

DocPereira/PEAL_V4_LHP_Zero_Entropy_Controlled is a source-linked Hugging Face repository assigned to Reinforcement learning. Its recorded task interface accepts environment state and produces action.

— context