Sami Haddadin -- The Great Robot Accelerator: Collective Learning of Optimal Embodied AI

In this talk

1. 1:37

   Why robotics needs to speed up

   Haddadin contrasts rapid progress in AI with the decade or two often needed to develop a new generation of robots. He argues that robotics must accelerate both intelligence and physical embodiment to keep pace.

2. 11:27

   Collective learning across robots

   For precise manipulation, Haddadin advocates combining physical models with data-driven methods. His examples include peg and key insertion, where tactile skills can transfer between tasks, and a network of 25 dual-arm robots that shared skills across more than 100 industrial components. In early trials, the system learned 100 skills in roughly two or three hours, while task order affected how quickly learning transferred.

3. 26:01

   AI-assisted robot design

   Haddadin proposes keeping human designers involved while AI helps create robot morphologies suited to specific tasks. He describes an experiment in which an evolutionary system designed a fingertip that could press a button it had not been explicitly told about; after about an hour, it was reliably activating the button. He also argues that physical testing can give design algorithms feedback that purely computational search misses.

4. 43:44

   Real-time optimal control for embodied intelligence

   Haddadin argues that motion planning must account for a robot’s internal dynamics to make use of compliance, elasticity, and energy storage. His team solves control problems offline, encodes the solutions, then generalizes them in real time; in their comparisons, dynamic movement primitives performed well for real-time trajectories. The approach was tested on reaching and catching tasks, with throwing as a further application.