Invited talk, Workshop on Reward Models for Large Language Models: Alignment, Capabilities, and Open Challenges, University of Oxford, Oxford, UK. 7 October 2026.
Vision-language-action (VLA) models give robots a strong prior, but a prior is not experience. This talk follows the path from preference data to on-robot experience: how preference and reward signals can be used to post-train VLA policies, why flow-based policies need finer-grained steps in online reinforcement learning, and how latency and hardware budgets shape which policy designs actually pay off on real platforms. I close with lessons from deploying reward-guided policies on a self-built mobile robot and on vehicle platforms.
This is an executable talk in the edtrace format used by Stanford CS336: the Python program is the slide deck. Text lines are the narration, code lines run and can be stepped through, and inspected values appear inline.
Powered by Jekyll and Minimal Light theme.