Position: Good Embodied Reward Models Need Bad Behavior Data

Thomas Tian, Yilin Wu, Andrea Bajcsy
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:172293-172305, 2026.

Abstract

This position paper argues that to obtain reliable embodied reward models, the community must invest in "bad" robot data: failed, suboptimal, error-prone, and even hazardous behaviors. While reward models are central to any foundation model’s lifecycle, today’s embodied reward models are trained primarily on successful behaviors. We analyze three state-of-the-art embodied reward models and find that they systematically over-reward behaviors that real human evaluators would penalize, including unsafe interactions, poor execution, and shortcut strategies that only superficially satisfy tasks. We attribute these failures to a key data gap: the scarcity of negative embodied data which is costly to collect and often filtered out or withheld in existing robotics datasets. Furthermore, we show that even modest exposure to real bad behavior data can improve alignment with human preferences and reduce costly false positives. We therefore call on the embodied AI community to curate and release their bad robot data, build synthetic bad data generation engines, develop more decentralized physical evaluation systems, and design benchmarks for fine-grained embodied reward model evaluations.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-tian26o, title = {Position: Good Embodied Reward Models Need Bad Behavior Data}, author = {Tian, Thomas and Wu, Yilin and Bajcsy, Andrea}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {172293--172305}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/tian26o/tian26o.pdf}, url = {https://proceedings.mlr.press/v306/tian26o.html}, abstract = {This position paper argues that to obtain reliable embodied reward models, the community must invest in "bad" robot data: failed, suboptimal, error-prone, and even hazardous behaviors. While reward models are central to any foundation model’s lifecycle, today’s embodied reward models are trained primarily on successful behaviors. We analyze three state-of-the-art embodied reward models and find that they systematically over-reward behaviors that real human evaluators would penalize, including unsafe interactions, poor execution, and shortcut strategies that only superficially satisfy tasks. We attribute these failures to a key data gap: the scarcity of negative embodied data which is costly to collect and often filtered out or withheld in existing robotics datasets. Furthermore, we show that even modest exposure to real bad behavior data can improve alignment with human preferences and reduce costly false positives. We therefore call on the embodied AI community to curate and release their bad robot data, build synthetic bad data generation engines, develop more decentralized physical evaluation systems, and design benchmarks for fine-grained embodied reward model evaluations.} }
Endnote
%0 Conference Paper %T Position: Good Embodied Reward Models Need Bad Behavior Data %A Thomas Tian %A Yilin Wu %A Andrea Bajcsy %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-tian26o %I PMLR %P 172293--172305 %U https://proceedings.mlr.press/v306/tian26o.html %V 306 %X This position paper argues that to obtain reliable embodied reward models, the community must invest in "bad" robot data: failed, suboptimal, error-prone, and even hazardous behaviors. While reward models are central to any foundation model’s lifecycle, today’s embodied reward models are trained primarily on successful behaviors. We analyze three state-of-the-art embodied reward models and find that they systematically over-reward behaviors that real human evaluators would penalize, including unsafe interactions, poor execution, and shortcut strategies that only superficially satisfy tasks. We attribute these failures to a key data gap: the scarcity of negative embodied data which is costly to collect and often filtered out or withheld in existing robotics datasets. Furthermore, we show that even modest exposure to real bad behavior data can improve alignment with human preferences and reduce costly false positives. We therefore call on the embodied AI community to curate and release their bad robot data, build synthetic bad data generation engines, develop more decentralized physical evaluation systems, and design benchmarks for fine-grained embodied reward model evaluations.
APA
Tian, T., Wu, Y. & Bajcsy, A.. (2026). Position: Good Embodied Reward Models Need Bad Behavior Data. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:172293-172305 Available from https://proceedings.mlr.press/v306/tian26o.html.

Related Material