TOPReward
Zero-shot reward model that reads the token log-probability of "True" from a frozen vision-language model asked whether a trajectory video completes a task instruction. No fine-tuned weights: the LeRobot port wraps Qwen3-VL (default Qwen3-VL-8B-Instruct) and returns one scalar log-prob reward per sample.
verified 2026-10-24paper →
Architecture
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 | topreward | docs |
| LeRobot: A reward model, not a policy: lives in src/lerobot/rewards/topreward, registered with RewardModelConfig.register_subclass("topreward") and selected as `reward_model.type=topreward` (not --policy.type). Inference-only and zero-shot; only the Qwen3-VL client is ported. Includes an offline labelling script that writes a SARM-compatible topreward_progress.parquet for RA-BC. | ||||
Checkpoints
- Qwen/Qwen3-VL-8B-Instruct · Default frozen VLM (vlm_name); TOPReward has no weights of its own — save_pretrained writes only config.json.
Related terms
Sources
- T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
TOPReward extracts token log-probabilities from an off-the-shelf VLM given a video and task instruction; zero-shot.
“TOPReward is a **zero-shot reward model** that extracts token log-probabilities from an off-the-shelf vision-language model (VLM) as a robotic reward signal.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/topreward.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
No fine-tuned weights to host; LeRobot wraps Qwen3VLForConditionalGeneration.
“Because the method only depends on a frozen VLM, TOPReward is **zero-shot**: there are no fine-tuned weights to host.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/topreward.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Registered as reward-model type `topreward`.
“@RewardModelConfig.register_subclass("topreward")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/rewards/topreward/configuration_topreward.py - T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
TOPReward was added in LeRobot v0.6.0.
“* feat(rewards): add TOPReward reward model by @s1lent4gnt in https://github.com/huggingface/lerobot/pull/3629”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0 - T3 · community reportedsource-codeverified 2026-10-24 (1d ago)
GitHub API reports the upstream TOPReward/TOPReward repo license as MIT.
https://github.com/TOPReward/TOPReward