Spring 2021 GRASP SFI: Anusha Nagabandi - February 17th, 2021
Anusha Nagabandi, a research scientist at Covariant and former UC Berkeley PhD researcher, presents work on model-based deep reinforcement learning for real-world robotics at the GRASP Student Faculty Industry Seminar hosted by GRASP Lab.
In this talk
- 1:37
Why real-world robots need to generalize
Nagabandi contrasts impressive demonstrations with failures caused by small position errors, uneven ground, and unseen factors. She argues that useful robots need to handle variation, such as moving different chairs across changing floors, rather than master one fixed setup.
- 8:35
Learning dynamics for control
She describes training a neural network to predict a robot’s next state from its current state and action, then using that model to evaluate candidate action sequences and replan. On the VelociRoACH six-legged robot, this approach produced controlled gaits after less than 20 minutes of training data. Adding camera images let one model handle multiple terrains without human labels.
- 17:14
Model-based learning for dexterous tasks
For dexterous manipulation, Nagabandi highlights learned models with ensembles to represent uncertainty and planning over candidate actions. A 24-joint hand learned tasks including object rotation and handwriting, with real-world rotations emerging after about four hours of training. The same model could support related motions, while model-free methods struggled when asked to follow arbitrary writing paths.
- 22:15
Fast adaptation without forgetting
Nagabandi explains how recent robot observations can update a model quickly when conditions change, such as a damaged leg or an added payload. She then describes maintaining a growing set of models for recurring tasks, using prediction error to decide which model fits and when to add another. This lets a robot improve on a returning task without losing performance after encountering a different one.
- 37:46
Inferring tasks and state from images
Nagabandi presents Melding, which uses image sequences, rewards, and a learned latent representation to infer both the robot’s state and its task. In a button-panel example, the robot explores to find the rewarded button, then reaches it directly in a later episode. She says future work could add senses such as touch and sound, and support planning over longer horizons.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.