Fall 2025 GRASP on Robotics - Jie Tan, Google DeepMind

In this talk

1. 2:29

   A decade of changing approaches

   Tan describes a goal of robots that can follow multimodal instructions, adapt to unfamiliar tasks and settings, and move with humanlike speed and dexterity. He traces the team’s shift from reinforcement learning and simulation to imitation learning and robot foundation models, citing RT-1 and the cross-institution RT-X dataset.

2. 9:07

   What enabled agile robots

   Tan credits stronger reinforcement learning methods, GPU simulation, and more affordable robot hardware with accelerating locomotion research. He says accurate actuator and latency models, plus domain randomization, helped transfer simulated quadruped policies to real machines.

3. 14:52

   Foundation models for robot skills

   Tan presents Gemini Robotics as a vision-language-action model that combines camera input, task instructions, and optional sensor readings to produce robot control targets. He highlights tests of visual, language, and action generalization, alongside examples such as selecting an unfamiliar toy for a basketball task, folding paper, and assembling a belt. Gemini Robotics 1.5 adds embodied reasoning, tool use, interleaved planning and action, and skill transfer across robot types; the earlier model reached about five hertz end-to-end control.

4. 34:16

   Data, hardware, and safety constraints

   Tan outlines a data pyramid: teleoperation offers useful labels but is costly, sensorized human demonstrations can have noisy action labels, simulation faces realism and world-building challenges, and internet data contributes broad prior knowledge. He also points to dexterous hands and tactile sensing as difficult hardware problems, and says safety remains hard because robots encounter many unusual situations near people.

5. 44:07

   Deployment and the road ahead

   Tan expects progress to come in uneven waves as new bottlenecks emerge, and estimates that superhuman robots may be more than a decade away, around 2035. In the Q&A, he says practical adoption depends on lowering cost and latency while improving success rates and safety; he also discusses cross-robot transfer, simulation, and the difficulty of reliably detecting task completion.