src/lerobot/templates/lerobot_rewardmodel_modelcard_template.md
{% if model_name == "reward_classifier" %} A reward classifier is a lightweight neural network that scores observations or trajectories for task success, providing a learned reward signal or offline evaluation when explicit rewards are unavailable. {% elif model_name == "sarm" %} A Success-Aware Reward Model (SARM) predicts a dense reward signal from observations, typically used downstream for reinforcement learning or human-in-the-loop fine-tuning when task success is not directly observable. {% elif model_name == "robometer" %} ROBOMETER is a general-purpose video-language robotic reward model built on a fine-tuned Qwen3-VL-4B backbone with progress, preference, and success heads. Given a trajectory video and a task description, it predicts dense, frame-level task progress in [0, 1] and frame-level success probabilities for downstream robot learning, including offline RL, online RL, data filtering and retrieval, and automated failure detection. {% elif model_name == "topreward" %} TOPReward is a zero-shot reward model that extracts token log-probabilities from an off-the-shelf vision-language model (default Qwen3-VL) as a reward signal. Given a video trajectory and a task instruction, it returns the VLM's log-likelihood of the instruction being true, with no fine-tuning required. {% else %} Reward model type not recognized — please update this template. {% endif %}
This reward model has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs.
lerobot-train \
--dataset.repo_id=${HF_USER}/<dataset> \
--reward_model.type={{ model_name | default("reward_classifier", true) }} \
--output_dir=outputs/train/<desired_reward_model_repo_id> \
--job_name=lerobot_reward_training \
--reward_model.device=cuda \
--reward_model.repo_id=${HF_USER}/<desired_reward_model_repo_id> \
--wandb.enable=true
Writes checkpoints to outputs/train/<desired_reward_model_repo_id>/checkpoints/.
from lerobot.rewards import make_reward_model
reward_model = make_reward_model(pretrained_path="<hf_user>/<reward_model_repo_id>")
reward = reward_model.compute_reward(batch)