RI Seminar: Abhishek Gupta : Simulation-Driven Robot Learning for the Lazy Researcher
Abhishek Gupta, professor at the University of Washington and director of the WEIRD Lab, presents at Carnegie Mellon’s Robotics Institute seminar on using simulation to improve robot learning with less task-specific engineering.
In this talk
- 1:32
From learned policies to ongoing improvement
Gupta contrasts model-and-planning-based robotics with learning from large datasets, then proposes systems that continue improving through deployment. He compares this direction with language-model training, where broad pretraining is followed by optimization for particular objectives.
- 10:41
Use simulation resets to make exploration easier
Gupta describes using simulator resets to break long tasks into shorter learning problems, with sampling focused on starting states where the policy has room to improve. This approach produced contact-heavy manipulation and assembly behaviors, though he notes that designing useful reset states still takes engineering.
- 21:17
Improve real-world policies by steering them in simulation
A policy that succeeds in simulation may rely on unrealistic contacts and fail on a physical robot. Gupta’s approach starts from a real-data policy and optimizes how a diffusion or flow policy is steered in simulation, keeping its actions within the base policy’s range. The method improved several tasks, but depends on broad behavioral coverage and can struggle with precise contacts.
- 35:08
Adapt to reality with short-horizon learning
Gupta argues that simulation can still help when its physics are imperfect: it can capture useful task structure even if its low-level contact predictions are wrong. His method keeps the simulated encoder and value function fixed, then updates the dynamics model from real robot experience; he reports gains using roughly a quarter to half an hour of physical data.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.