RSS 2026 Day 4 Part 1

In this talk

1. 7:21

   Perception and scene understanding

   Presenters describe methods for open-vocabulary occupancy mapping, tactile imaging, anonymous robot localization, event-camera motion segmentation, and template-guided object detection. Examples include reconstructing hidden inclusions from touch and detecting target objects in clutter using local visual matches.

2. 28:55

   Efficient representations and safer reconstruction

   Talks cover tetrahedron-based signed distance fields for accurate, deformable collision detection, plus control methods that account for perception and dynamics uncertainty. Scene reconstruction methods use physical constraints to prevent objects from floating or penetrating one another, with examples tested in simulation and on robots.

3. 49:46

   Learning from motion, touch, and contact

   Presenters plan views of moving objects by accounting for uncertainty in where objects will be when a robot arrives, and describe soft grippers that compensate for force loss over time. Wenzhen Yuan traces tactile manipulation from force and slip feedback to learned representations, sensor simulation, and sensor designs intended to transfer across devices.

4. 2:07:37

   Planning under uncertainty and physical constraints

   The planning session includes vectorized POMDP search, prepared recovery branches for manipulation failures, and hypothesis-based world model updates for unfamiliar environments. Other work uses uncertainty-aware world models, informative sensing paths, and the Kinder benchmark to study physical reasoning.

5. 2:33:25

   Fast optimization and certified control

   Presenters introduce hierarchical inverse kinematics, horizon-optimized trajectory planning, reusable representations of redundant robot configurations, and planning for steerable or continuum robots. Reachability methods bound possible outcomes under uncertainty, with demonstrations including contact-rich manipulation and real-world control.