Haruki Nishimura, a Stanford PhD researcher in the Multi-Robot Systems Lab, presents his thesis on planning robot trajectories under uncertainty in interactive environments, with advisor Mac Schwager and the defense committee at Stanford.
Nishimura contrasts robots in controlled warehouses with machines operating around people and unpredictable obstacles. His first topic is active perception: robots can choose actions that improve what they know, such as a manipulation robot learning an object's physical properties while moving it. He introduces stochastic sequential action control, which uses sampled future observations and a fast optimization step to adjust a nominal policy.
Nishimura turns to uncertainty in how people will move, including rare events that could cause serious collisions. His risk-sensitive planner combines sequential action control with Trajectron++ pedestrian forecasts, which represent multiple plausible paths. In a five-person experiment, the risk-neutral robot collided, while the risk-sensitive version completed five runs without a collision and kept more distance from pedestrians.
Nishimura shows that a planner can fail when its model treats people as static obstacles with Gaussian random motion. He proposes distributionally robust control, which plans against a set of plausible models defined using KL divergence. The RAT-iLQR method adjusts its risk level automatically; in a collision-avoidance simulation it avoided collisions and reduced average trajectory-tracking error by 16% compared with a maximally risk-sensitive setting.
Nishimura says his work focuses on planning and control, while modern autonomy systems also rely on learned perception and prediction. He argues these components need a shared interface so uncertainty can be handled across the full system. In the Q&A, he describes machine learning as useful for building predictive models and estimating their errors, which could help provide safety assurances.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Mac Schwager, a robotics researcher speaking for his lab at MIT Robotics, examines whether vision-language-action policies generalize and what might make them more reliable.
