Can a single reward model work zero-shot across robots, tasks, and environments? We’re excited to introduce Robometer, a general-purpose robotic reward model trained on 1M+ trajectories that enables: - Online RL - Offline RL - Model-based RL - Data retrieval + imitation learning - Automatic failure detection And more—all zero-shot across robots, tasks, and scenes. Most existing reward models are trained only on successful or reward-labeled demonstrations, which limits their ability to distinguish between successful and failed attempts. Robometer addresses this by using trajectory comparisons, enabling reward models to learn from reward-unlabeled and failure trajectories. Robometer is a 4B-parameter video-language reward model predicting:  - Dense reward (0-1 progress)  - Success  - Trajectory preferences (which attempt was better) We train by pairing input trajectories of different expertise levels and tasks, and by using video rewind. Robometer’s ability to train on reward-unlabeled, failed robot trajectories allows us to scale to training on our RBM-1M dataset: - 21 embodiments: bimanual, humanoid, mobile, tabletop - 1M+ trajs - 140k+ trajectories from mixed-expertise datasets We evaluate Robometer both as a reward estimator and as a component for robot learning. On self-collected, unseen data from three universities, Robometer outperforms existing reward models. We further evaluate Robometer in real-world experiments across four universities (USC, MIT, UTD, UW) on tasks including online RL, offline RL, model-based RL, data retrieval for imitation learning, and zero-shot failure detection. Across these settings, Robometer consistently improves robot learning performance compared to using prior reward models. We hope Robometer can help accelerate research on general-purpose robot learning. We encourage you to try out the model yourself for your own experiments zero-shot or fine-tune it on your own data! Project: http://robometer.github.io Paper: https://lnkd.in/gz7A78bD Code: https://lnkd.in/gdNh2Qej This project was co-led with Anthony Liang. Huge thanks to all our incredible collaborators + advisors: Jiahui Zhang, Minyoung Hwang, Abrar A., Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Bıyık, and Jesse Zhang.

Can a single reward model work zero-shot across robots, tasks, and environments? We’re excited to introduce Robometer, a general-purpose robotic reward model trained on 1M+ trajectories that enables: - Online RL - Offline RL - Model-based RL - Data retrieval + imitation learning - Automatic failure detection And more—all zero-shot across robots, tasks, and scenes. Most existing reward models are trained only on successful or reward-labeled demonstrations, which limits their ability to distinguish between successful and failed attempts. Robometer addresses this by using trajectory comparisons, enabling reward models to learn from reward-unlabeled and failure trajectories. Robometer is a 4B-parameter video-language reward model predicting:  - Dense reward (0-1 progress)  - Success  - Trajectory preferences (which attempt was better) We train by pairing input trajectories of different expertise levels and tasks, and by using video rewind. Robometer’s ability to train on reward-unlabeled, failed robot trajectories allows us to scale to training on our RBM-1M dataset: - 21 embodiments: bimanual, humanoid, mobile, tabletop - 1M+ trajs - 140k+ trajectories from mixed-expertise datasets We evaluate Robometer both as a reward estimator and as a component for robot learning. On self-collected, unseen data from three universities, Robometer outperforms existing reward models. We further evaluate Robometer in real-world experiments across four universities (USC, MIT, UTD, UW) on tasks including online RL, offline RL, model-based RL, data retrieval for imitation learning, and zero-shot failure detection. Across these settings, Robometer consistently improves robot learning performance compared to using prior reward models. We hope Robometer can help accelerate research on general-purpose robot learning. We encourage you to try out the model yourself for your own experiments zero-shot or fine-tune it on your own data! Project: http://robometer.github.io Paper: https://lnkd.in/gz7A78bD Code: https://lnkd.in/gdNh2Qej This project was co-led with Anthony Liang. Huge thanks to all our incredible collaborators + advisors: Jiahui Zhang, Minyoung Hwang, Abrar A., Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Bıyık, and Jesse Zhang.

Can a single reward model work zero-shot across robots, tasks, and environments? We’re excited to introduce Robometer, a general-purpose robotic reward model trained on 1M+ trajectories that enables: - Online RL - Offline RL - Model-based RL - Data retrieval + imitation learning - Automatic failure detection And more—all zero-shot across robots, tasks, and scenes. Most existing reward models are trained only on successful or reward-labeled demonstrations, which limits their ability to distinguish between successful and failed attempts. Robometer addresses this by using trajectory comparisons, enabling reward models to learn from reward-unlabeled and failure trajectories. Robometer is a 4B-parameter video-language reward model predicting:  - Dense reward (0-1 progress)  - Success  - Trajectory preferences (which attempt was better) We train by pairing input trajectories of different expertise levels and tasks, and by using video rewind. Robometer’s ability to train on reward-unlabeled, failed robot trajectories allows us to scale to training on our RBM-1M dataset: - 21 embodiments: bimanual, humanoid, mobile, tabletop - 1M+ trajs - 140k+ trajectories from mixed-expertise datasets We evaluate Robometer both as a reward estimator and as a component for robot learning. On self-collected, unseen data from three universities, Robometer outperforms existing reward models. We further evaluate Robometer in real-world experiments across four universities (USC, MIT, UTD, UW) on tasks including online RL, offline RL, model-based RL, data retrieval for imitation learning, and zero-shot failure detection. Across these settings, Robometer consistently improves robot learning performance compared to using prior reward models. We hope Robometer can help accelerate research on general-purpose robot learning. We encourage you to try out the model yourself for your own experiments zero-shot or fine-tune it on your own data! Project: http://robometer.github.io Paper: https://lnkd.in/gz7A78bD Code: https://lnkd.in/gdNh2Qej This project was co-led with Anthony Liang. Huge thanks to all our incredible collaborators + advisors: Jiahui Zhang, Minyoung Hwang, Abrar A., Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Bıyık, and Jesse Zhang.

Can a single reward model work zero-shot across robots, tasks, and environments? We’re excited to introduce Robometer, a general-purpose robotic reward model trained on 1M+ trajectories that enables: - Online RL - Offline RL - Model-based RL - Data retrieval + imitation learning - Automatic failure detection And more—all zero-shot across robots, tasks, and scenes. Most existing reward models are trained only on successful or reward-labeled demonstrations, which limits their ability to distinguish between successful and failed attempts. Robometer addresses this by using trajectory comparisons, enabling reward models to learn from reward-unlabeled and failure trajectories. Robometer is a 4B-parameter video-language reward model predicting:  - Dense reward (0-1 progress)  - Success  - Trajectory preferences (which attempt was better) We train by pairing input trajectories of different expertise levels and tasks, and by using video rewind. Robometer’s ability to train on reward-unlabeled, failed robot trajectories allows us to scale to training on our RBM-1M dataset: - 21 embodiments: bimanual, humanoid, mobile, tabletop - 1M+ trajs - 140k+ trajectories from mixed-expertise datasets We evaluate Robometer both as a reward estimator and as a component for robot learning. On self-collected, unseen data from three universities, Robometer outperforms existing reward models. We further evaluate Robometer in real-world experiments across four universities (USC, MIT, UTD, UW) on tasks including online RL, offline RL, model-based RL, data retrieval for imitation learning, and zero-shot failure detection. Across these settings, Robometer consistently improves robot learning performance compared to using prior reward models. We hope Robometer can help accelerate research on general-purpose robot learning. We encourage you to try out the model yourself for your own experiments zero-shot or fine-tune it on your own data! Project: http://robometer.github.io Paper: https://lnkd.in/gz7A78bD Code: https://lnkd.in/gdNh2Qej This project was co-led with Anthony Liang. Huge thanks to all our incredible collaborators + advisors: Jiahui Zhang, Minyoung Hwang, Abrar A., Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Bıyık, and Jesse Zhang.

Can a single reward model work zero-shot across robots, tasks, and environments? We’re excited to introduce Robometer, a general-purpose robotic reward model trained on 1M+ trajectories that enables: - Online RL - Offline RL - Model-based RL - Data retrieval + imitation learning - Automatic failure detection And more—all zero-shot across robots, tasks, and scenes. Most existing reward models are trained only on successful or reward-labeled demonstrations, which limits their ability to distinguish between successful and failed attempts. Robometer addresses this by using trajectory comparisons, enabling reward models to learn from reward-unlabeled and failure trajectories. Robometer is a 4B-parameter video-language reward model predicting:  - Dense reward (0-1 progress)  - Success  - Trajectory preferences (which attempt was better) We train by pairing input trajectories of different expertise levels and tasks, and by using video rewind. Robometer’s ability to train on reward-unlabeled, failed robot trajectories allows us to scale to training on our RBM-1M dataset: - 21 embodiments: bimanual, humanoid, mobile, tabletop - 1M+ trajs - 140k+ trajectories from mixed-expertise datasets We evaluate Robometer both as a reward estimator and as a component for robot learning. On self-collected, unseen data from three universities, Robometer outperforms existing reward models. We further evaluate Robometer in real-world experiments across four universities (USC, MIT, UTD, UW) on tasks including online RL, offline RL, model-based RL, data retrieval for imitation learning, and zero-shot failure detection. Across these settings, Robometer consistently improves robot learning performance compared to using prior reward models. We hope Robometer can help accelerate research on general-purpose robot learning. We encourage you to try out the model yourself for your own experiments zero-shot or fine-tune it on your own data! Project: http://robometer.github.io Paper: https://lnkd.in/gz7A78bD Code: https://lnkd.in/gdNh2Qej This project was co-led with Anthony Liang. Huge thanks to all our incredible collaborators + advisors: Jiahui Zhang, Minyoung Hwang, Abrar A., Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Bıyık, and Jesse Zhang.

Can a single reward model work zero-shot across robots, tasks, and environments? We’re excited to introduce Robometer, a general-purpose robotic reward model trained on 1M+ trajectories that enables: - Online RL - Offline RL - Model-based RL - Data retrieval + imitation learning - Automatic failure detection And more—all zero-shot across robots, tasks, and scenes. Most existing reward models are trained only on successful or reward-labeled demonstrations, which limits their ability to distinguish between successful and failed attempts. Robometer addresses this by using trajectory comparisons, enabling reward models to learn from reward-unlabeled and failure trajectories. Robometer is a 4B-parameter video-language reward model predicting:  - Dense reward (0-1 progress)  - Success  - Trajectory preferences (which attempt was better) We train by pairing input trajectories of different expertise levels and tasks, and by using video rewind. Robometer’s ability to train on reward-unlabeled, failed robot trajectories allows us to scale to training on our RBM-1M dataset: - 21 embodiments: bimanual, humanoid, mobile, tabletop - 1M+ trajs - 140k+ trajectories from mixed-expertise datasets We evaluate Robometer both as a reward estimator and as a component for robot learning. On self-collected, unseen data from three universities, Robometer outperforms existing reward models. We further evaluate Robometer in real-world experiments across four universities (USC, MIT, UTD, UW) on tasks including online RL, offline RL, model-based RL, data retrieval for imitation learning, and zero-shot failure detection. Across these settings, Robometer consistently improves robot learning performance compared to using prior reward models. We hope Robometer can help accelerate research on general-purpose robot learning. We encourage you to try out the model yourself for your own experiments zero-shot or fine-tune it on your own data! Project: http://robometer.github.io Paper: https://lnkd.in/gz7A78bD Code: https://lnkd.in/gdNh2Qej This project was co-led with Anthony Liang. Huge thanks to all our incredible collaborators + advisors: Jiahui Zhang, Minyoung Hwang, Abrar A., Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Bıyık, and Jesse Zhang.