Matrix/Policy·Reward model·Anthony Liang, Jesse Zhang et al. (University of Southern California, University of Washington, Ai2, NVIDIA and others)

Robometer

General-purpose video-language robotic reward model on a Qwen3-VL-4B backbone with progress, success and preference heads. Predicts dense per-frame task progress and success from a trajectory video and task description; in LeRobot it is used zero-shot (inference-only) and returns one scalar reward per sample.

open weightsverified 2026-10-24paper →

Reward models are not policies you deploy on a robot, so they are listed separately from the matrix.

Architecture

Family: reward-model
Backbone: Qwen3-VL-4B-Instruct + progress / success / preference headssrc
Action head: scalar-rewardsrc
Code license: Apache-2.0 (LeRobot integration)
Weights license: Apache-2.0
Cameras: min 1 · max 1src
Language input: required

Compute requirements

No sourced requirements yet.

Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.

Measured inference 0 runs · inference cost only, not task success

No measured runs yet. Submit one with its runtime, precision and camera count.

Framework support

Framework	Support	Since	Type string	Docs
LeRobot	nativesrc	0.6.0	robometer	docs
LeRobot: A reward model, not a policy: lives in src/lerobot/rewards/robometer, registered with RewardModelConfig.register_subclass("robometer") and selected as `reward_model.type=robometer` (not --policy.type). Inference-only in LeRobot (no training or preference-pair training). Ships a compute_rabc_weights module for offline progress labelling.

Checkpoints

lerobot/Robometer-4B · Published ROBOMETER checkpoint used by the LeRobot integration (Hub license tag apache-2.0).

Related terms

Reward model

Sources

T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
ROBOMETER builds on Qwen3-VL-4B with progress, success and preference heads; LeRobot integration is inference-only.
“ROBOMETER builds on `Qwen/Qwen3-VL-4B-Instruct` and adds three lightweight prediction heads:”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/robometer.mdx
T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Video-language only; no privileged robot state.
“ROBOMETER is video-language based; it does not use privileged robot state such as contact forces or object poses.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/robometer.mdx
T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot integration is inference-only.
“The LeRobot integration is currently **inference-only**.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/robometer.mdx
T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Registered as reward-model type `robometer`.
“@RewardModelConfig.register_subclass("robometer")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/rewards/robometer/configuration_robometer.py
T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
ROBOMETER was added in LeRobot v0.6.0.
“* feat(rewards): add ROBOMETER reward model by @s1lent4gnt in https://github.com/huggingface/lerobot/pull/3627”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0
T3 · community reportedhub-checkpointverified 2026-10-24 (1d ago)
Hub API: lerobot/Robometer-4B exists, tagged license:apache-2.0.
https://huggingface.co/lerobot/Robometer-4B