ICRA 2026 Keynote 5 - Robot Learning & Models

In this talk

1. 0:51

   Structure helps robots handle unfamiliar tasks

   David Hu asks whether robots can make decisions in unfamiliar homes, where appliances and their controls may be new. He shows that extracting states, actions, and transition rules from appliance manuals helps a robot set devices such as an induction cooker more reliably. For broader household tasks, robots can track uncertain hypotheses, test them in the environment, and replan when they are wrong.

2. 21:45

   Programming for reliable robot skills

   Stephanie Tellex argues that robots will still need a way for people to specify tasks, despite progress in large models. She describes combining learning from demonstration with reinforcement learning, and compares formal task representations with end-to-end language-to-action models. Her examples include using structured skills to improve object grasping with fewer demonstrations and generalize across poses and robot platforms.

3. 42:45

   Geometry as an inductive bias

   Noémie Le Lar shows how representing robot variables in their natural geometric spaces can improve learning. Riemannian flow matching handles large orientation changes, while control theory adds stability guarantees to learned motion policies. For deformable objects, geometric Lagrangian models compress a cloth with 600 degrees of freedom into a 10-dimensional latent space while preserving physical structure.

4. 1:04:04

   Planning around model uncertainty

   Paolo Robuffo Giordano presents a sensitivity matrix that estimates how model parameter errors affect a robot's motion under its existing controller. He uses it to optimize trajectories and build uncertainty tubes for constraints, with examples including drones passing through windows and avoiding motor saturation. The same sensitivity measure can help select robust learned policies and is being explored as a reinforcement learning reward.