LeRobot TOPReward

Community implementation

Zero-shot reward wrapper using an off-the-shelf VLM token probability.

View full model description

Research evaluation and reward estimation for robot trajectories.

Papers & reports (1)

TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics

Paper available

arXiv: 2602.19313DOI: 10.48550/arxiv.2602.19313
Authors & affiliations (9)

Author order: Original order

  1. Shirui ChenCorresponding author

    University of Washington · Allen Institute for AI

  2. Cole Harrison

    Amazon

  3. Ying-Chun Lee

    University of Washington

  4. Angela Jin Yang

    University of Washington

  5. Zhongzheng Ren

    University of Washington · Allen Institute for AI · University of North Carolina at Chapel Hill

  6. Lillian J. Ratliff

    University of Washington

  7. Jiafei DuanCo-advising

    University of Washington · Allen Institute for AI

  8. Dieter FoxCo-advising

    University of Washington · Allen Institute for AI

  9. Ranjay KrishnaCo-advising

    University of Washington · Allen Institute for AI

First published
2026-02-22
Latest paper revision
v2 · 2026-07-22
Venue
Not provided

Covered model versions

  • topreward · LeRobot TOPReward

    Original method paper referenced by the LeRobot implementation; not a separate evaluation of every LeRobot integration change.

Paper ↗PDF ↗

How to access and use

Public code resources
Public access

Training / fine-tuning code

Access
Not applicable

Verified on · 2026-10-26

Release information

SourceVerified on · 2026-10-23

Checks before deployment

Estimated rewards depend on the observation and task distribution and must not be treated as a safety guarantee. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Runtime and frameworks
LeRobot / Transformers
Fine-tuning requirements
The LeRobot port is inference-only and zero-shot; it uses a frozen Qwen3-VL model and does not train a reward model.

Architecture and inputs / outputs

Model type
Reward & evaluation
Backbone
Qwen3-VL-8B-Instruct by default
Input
Language / RGB images

Scope and training

Languages

English

Functional purpose

Reward & progress

Record identity

Community implementation

Pipeline task

Robot reward modeling

Prediction output

Token log-probability reward; optional binary threshold

Related robots

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Publisher evaluation · Reported; not independently reproduced

TOPReward Qwen3-VL-8B paper reference · OXE:33 filtered datasets ↗

VOC: 0.874

Original TOPReward v2 paper with Qwen3-VL-8B and direct prompt scoring. VOC is temporal value-order rank correlation on[-1,1], not robot task success or calibrated completion. This is an author-reported method reference, not an independently reproduced LeRobot integration run or a standalone trained reward checkpoint. Table1: mean dataset VOC over33 filtered OXE datasets,20 episodes each. The project page’s39-dataset0.857 is a different published scope.

Publisher evaluation · Reported; not independently reproduced

TOPReward Qwen3-VL-8B paper reference · ManiRewardBench: successful subset ↗

VOC: 0.942

Original TOPReward v2 paper with Qwen3-VL-8B and direct prompt scoring. VOC is temporal value-order rank correlation on[-1,1], not robot task success or calibrated completion. This is an author-reported method reference, not an independently reproduced LeRobot integration run or a standalone trained reward checkpoint. Table2: metrics averaged over497 successful-evaluation episodes,113 tasks across four platforms. The full benchmark has130 unique tasks including a failure split. Not the project headline0.947/0.945, and not a success-detection AUC.

Full parameters and verification notes (model card)

Original paper: TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics

First submitted: 2026-02-22 (arXiv:2602.19313v1).

First and corresponding author: Shirui Chen (sc256@uw.edu).

Co-advisors: Jiafei Duan, Dieter Fox, Ranjay Krishna.

LeRobot integration maintainer: Hugging Face. LeRobot.

The LeRobot wrapper is inference-only: a frozen Qwen3-VL supplies token log-probabilities, with no separate task-trained reward weights. Raw reward is a log-probability, not a calibrated completion percentage; an optional finite threshold returns a binary value. Per-episode min/max progress normalization uses the full sampled trajectory and is non-causal, separate from online raw reward. Only Qwen3-VL clients are supported by this port. Original TOPReward code is MIT; LeRobot code is Apache2 and backbone weights retain their own terms. Exact LeRobot release date and hosted API remain unverified.

Submit a source or correction