aliangdw / Robometer-4B-LIBERO-No-Fail

Robometer 4B LIBERO

Paper: arXiv (Coming Soon)

Robometer is a general-purpose vision-language reward model for robotics. It is trained on RBM-1M with Qwen3-VL-4B to predict per-frame progress , per-frame success , and trajectory preferences from rollout videos. The model combines (1) frame-level progress supervision on expert data and (2) trajectory-comparison preference supervision, so it can learn from both successful and failed rollouts and generalize across diverse robot embodiments and tasks.

Given a task instruction and a rollout video (or frame sequence), the model predicts:

• Per-frame progress — continuous progress values over time (e.g. 0–1 or binned).
• Per-frame success — success probability (or binary) at each timestep.
• Preference / ranking — which of two trajectories is better for the task.

This model is trained on LIBERO-10/Spatial/Object/Goal only on preference and progress objectives but without the failure dataset.

Usage

For full setup, example scripts, and configs, see the GitHub repo : github.com/aliang8/robometer .

Option 1 — Run the model locally (loads this checkpoint from Hugging Face):

uv run python scripts/example_inference_local.py \
  --model-path aliangdw/Robometer-4B-LIBERO \
  --video /path/to/video.mp4 \
  --task "your task description"

Option 2 — Use the evaluation server (start server, then run client):

# Start server
uv run python robometer/evals/eval_server.py \
  --config-path=robometer/configs \
  --config-name=eval_config_server \
  model_path=aliangdw/Robometer-4B-LIBERO \
  server_url=0.0.0.0 \
  server_port=8000

# Client (no robometer dependency)
uv run python scripts/example_inference.py \
  --eval-server-url http://localhost:8000 \
  --video /path/to/video.mp4 \
  --task "your task description"

Citation

If you use this model, please cite:

@misc{robometer2025,
  title={Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons},
  author={Anthony Liang* and Yigit Korkmaz* and Jiahui Zhang and Minyoung Hwang and Abrar Anwar and Sidhant Kaushik and Aditya Shah and Alex S. Huang and Luke Zettlemoyer and Dieter Fox and Yu Xiang and Anqi Li and Andreea Bobu and Abhishek Gupta and Stephen Tu† and Erdem B{\i}y{\i}k† and Jesse Zhang†},
  year={2025},
  url={https://github.com/aliang8/reward_fm},
  note={arXiv coming soon}
}