Learning Robotic Locomotion Affordances and Photorealistic Simulators from Human-Captured Data

Abstract

Learning reliable affordance models which satisfy human preferences is often hindered by a lack of high-quality training data. Similarly, learning visuomotor policies in simulation can be challenging due to the high cost of photo-realistic rendering. We present PAWS: a comprehensive robot learning framework that uses a novel portable data capture rig and processing pipeline to collect long-horizon trajectories that include camera poses, foot poses, terrain meshes, and 3D radiance fields. We also contribute PAWS-Data: an extensive dataset gathered with PAWS containing over 10 hours of indoor and outdoor trajectories spanning a variety of scenes. With PAWS-Data we leverage radiance fields' photo-realistic rendering to generate tens of thousands of viewpoint-augmented images, then produce pixel affordance labels by identifying semantically similar regions to those traversed by the user. On this data we finetune a navigation affordance model from a pretrained backbone, and perform detailed ablations. Additionally, We open source PAWS-Sim, a high-speed photo-realistic simulator which integrates PAWS-Data with IsaacSim, enabling research for visuomotor policy learning. We evaluate the utility of the affordance model on a quadrupedal robot, which plans through affordances to follow pathways and sidewalks, and avoid human collisions. Project resources are available on the [website](https://pawslocomotion.com).