2025 Winter Robotics Colloquium: Sebastian Castro (Robotics and AI Institute)

In this talk

1. 8:14

   Why planning needs execution support

   Castro introduces task planning as searching for a sequence of actions that can change the world from its current state to a goal, such as moving an apple onto a table. He explains that abstract plans may fail when the robot lacks room to place an object or when the world differs from its model. Task and motion planning, reactive skills, and human input can help address those gaps.

2. 12:49

   Connecting plans to robot behavior

   Castro describes using modular robot software, behavior trees, and world models to connect planner actions with controllers and feed observed changes back into planning. In a CSAIL example, a robot unable to reach an apple first moves a green can out of the way, updates object locations, and retries. A PickNik project used PlanSys2 and MoveIt Task Constructor for cargo handling, with options for operator recovery and motion approval.

3. 32:16

   Learning skills that can be composed

   Castro presents work at the Robotics and AI Institute on learned policies, including an equivariant diffusion policy trained on about 60 demonstrations for a bagel task. He identifies three open problems: recognizing when a skill has finished, combining skills in new sequences, and handling changes in the environment. Predicting symbolic conditions could help planners decide whether to continue, recover, or replan.

4. 43:40

   Keep systems modular and inspectable

   Castro discusses ways to improve skill recombination, including adding motion planning between learned actions, using object-relative representations, and conditioning policies on task context. He argues that learned methods should work alongside established tools such as inverse kinematics, state estimation, and symbolic planning. He also emphasizes logging, visualization, and accessible communication as practical parts of robotics work.