RI Seminar: Roberto Martín-Martín : Expanding What Robots Can Learn in the Real World
Roberto Martín-Martín, professor at UT Austin and head of the RPL Lab, explains at the CMU Robotics Institute seminar how robots can build new real-world skills by reusing prior learning, human guidance, and carefully chosen data.
In this talk
- 1:25
Build skills before a task arrives
Martín-Martín draws on infant development: exploratory banging can lay groundwork for learning to hammer. In a robot experiment, policies that learned to bang first acquired targeted hammering faster than learning from scratch, though a fixed hierarchical policy plateaued because it could not refine the low-level skill.
- 9:11
Shape exploration and learn safely
SLAC learns a latent action space in simulation, using unsupervised skill discovery and safety rewards to give a mobile manipulator safer, more useful motions to explore in the real world. Across four tasks, including wiping and sweeping, policies trained in under an hour; SLAC reached high success and recorded few safety violations. Safe Mimic applies related safety estimates to retarget a single human video demonstration to a robot, reporting 40% to 100% success across seven tasks.
- 18:12
Extract more from human guidance
Martín-Martín argues that demonstrations and corrections contain information beyond the actions themselves, including intent and reasons for intervention. Safe Mimic uses video-derived task phases to guide safe adaptation, while a VLA correction method turns verbal explanations into gated costs that can steer π0.5 with five to ten corrections and generalize across scenes and objects.
- 34:19
Choose data that improves the policy
Data models estimate which training examples will help a target policy, allowing Data Mill to select useful trajectories from larger datasets such as OXE. The lab also uses data attribution to guide where people should collect additional examples; in reported experiments, this approach outperformed existing guidance methods and human experts.
- 45:19
Keep learning without forgetting
The lab explores memory for carrying useful experience across buildings and appliances, as well as continual reinforcement learning for pretrained VLAs. A recipe combining parameter-efficient fine-tuning with online, on-policy reinforcement learning reached 88% success against a 91% oracle, with near-zero forgetting and strong performance on held-out tasks.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.