Robot
SO-101
Robometer 4BReal-robot experiment
Robometer's official results include real-world evaluation on an SO-101 platform.
Applies to the reported experimental setup and data collection protocol.
Robometer 4B
Pretrained reward model scoring task progress and success from video and language.
Research evaluation and scoring of robot behavior with Robometer.
Access and availability
- Code
- Public access
- Model files
- Public access
Code and training / fine-tuning
- Access
- Public access
Verified on · 2026-10-27
- Code license
- MIT
Model files
| File | Access | Weights license |
|---|---|---|
| robometer/Robometer-4B | Access |
Getting started
- Runtime and frameworks
- LeRobot / TransformersNVIDIA GPU · bfloat16
- Deployment requirements
- Python 3.10+, NVIDIA GPU drivers; inference HTTP server is self-hosted.
At a glance
- Model type
- Reward model
- Backbone
- Qwen3-VL-4B
- Weight parameters
- 4.4503 B
- Input
- Language instruction / RGB camera images
- Released
- March 2, 2026
Related robots
Evaluation results
No verified independent benchmark results are recorded for this version or the selected weights.
Publisher evaluation · Reported; not independently reproduced
RBM-EVAL · Out-of-distribution ↗VOC Pearson r: 0.94
Full RBM-1M training; out-of-distribution reward evaluation. Authors report rounded aggregate results.
Publisher evaluation · Reported; not independently reproduced
RBM-EVAL · Out-of-distribution ↗Trajectory ranking: Kendall tau: 0.64
Full RBM-1M training; out-of-distribution reward evaluation. Authors report rounded aggregate results.
Technical details
Languages
English
Functional purpose
Reward & progress
Pipeline task
Robot reward modeling
Prediction output
Task progress and success score
Resources
Model card
Paper: Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons. arXiv:2603.02115, first submitted 2026-03-02. Anthony Liang and Yigit Korkmaz contributed equally as first authors. Stephen Tu, Erdem Bıyık and Jesse Zhang provided equal advising. The paper does not designate a corresponding author. [arXiv](https://arxiv.org/abs/2603.02115) · [PDF v1](https://arxiv.org/pdf/2603.02115v1) · [Project](https://robometer.github.io/)
Before you use this version: Reward scores are model judgments and can be sensitive to camera view, task wording and distribution shift. They are not a physical safety signal. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.
Submit a source or correction
