Fall 2025 GRASP SFI - Roberto Martín Martín, University of Texas at Austin

In this talk

1. 0:40

   Why home robots need mobile manipulation

   Martín argues that useful household robots must navigate and manipulate objects, since many desired tasks involve both. A survey of thousands of people found that 92% of their top 10,000 requested tasks were household activities. These long-horizon tasks also involve contact and force, and must work across varied homes.

2. 6:46

   Adapting one human demonstration safely

   SafeMimic uses a single human video to identify task phases, then lets a robot adapt the demonstrated actions to its own body. Safety value functions trained in simulation flag risks such as excessive force or dropping an object; the robot can backtrack and try other actions. In seven real-world tasks, the method succeeded using one video in the same environment and with the same objects, while direct motion retargeting largely failed.

3. 29:09

   Multiplying demonstrations with simulation

   MiniGen turns one simulated demonstration into many variations by changing navigation and robot poses while preserving the contact-heavy manipulation steps. Policies trained on this data were more robust to different starting positions. In a cup-grasping test, simulated pretraining raised success from 0% to 60% when both approaches were fine-tuned with 50 real-world demonstrations.

4. 39:52

   Selecting useful data for a task

   DataMill uses data models to estimate which subsets of a large dataset will help a particular imitation-learning task. Instead of repeatedly testing policies in rollouts, it uses action prediction on a held-out validation set as a cheaper proxy. In one experiment, a few Tiago demonstrations guided selection from OXE, which contains no data from Tiago, to train a policy for that robot.

5. 56:35

   Simulation as a source of safety priors

   Martín describes learning a safe action space in simulation through unsupervised skill discovery, then using it to train downstream tasks with reinforcement learning in the real world. The real-world training took less than an hour in the examples shown. He says transferring a safety prior can be easier than transferring a complete policy because the policy learning still happens on the robot.