Toyota Research Institute (TRI) is developing general-purpose robots capable of performing a wide variety of dexterous tasks by building general-purpose machine learning foundation models for dexterous robot manipulation, referred to as Large Behavior Models (LBMs). These LBMs utilize generative AI techniques to produce robot actions from sensor data and human requests, requiring large datasets of embodied robot demonstrations, internet-scale text, image, and video data, and high-quality simulation.
Requirements
- Hands-on experience with using machine learning for learned control, including RL, offline RL or behavior cloning, for manipulation. Or experience with machine learning and familiarity with large multi-modal datasets and models.
- Strong software development skills in Python.
- Experience working with large-scale datasets and multi-node training
Responsibilities
- Data-efficient and general algorithms for learning robust policies using multiple sensing modalities: proprioception, images, 3D representations, force, and dense tactile sensing.
- Scaling learning approaches to large-scale models trained on diverse sources of data, including web-scale text, images, and video.
- Leveraging test time computation for embodied applications.
- Quick and efficient improvement of learned policies.
- Continual Learning and Adaption
- Multi-Modal Reasoning Models.
- Structured hierarchical reasoning using learned models.
Other
- A “make it happen” demeanour and comfort with fast prototyping.
- A passion for robotics and doing research grounded in important fundamental problems.
- A track record of relevant publications in top international conferences (RSS, NeuRIPS, ICML, ICLR, CoRL, ICRA, IROS, …)
- Hardware experience.
- A track record of relevant open-source software contributions
