When a pedestrian blocks a robot's path, its intended movement and the motion it can actually make may differ. A Lehigh University paper posted Sept. 30 and revised Oct. 1 presents Social-WM, a planner that uses this mismatch as a warning while plotting a route through shared indoor spaces.
During training, Social-WM learns from observed scene changes that a command to move forward can end in a stop or a deviation when the path is blocked. At planning time, it imagines several short movement sequences, predicts the scene each might lead to and estimates which motion would be realizable. A large gap between intended and expected motion warns that an option may conflict with people or obstacles. The planner can choose another candidate, carries out only the first part of its plan and observes again. It does not explicitly track pedestrians or use privileged measurements of their state.
In the paper's Habitat simulation benchmark, Social-HM3D, the planner reached its navigation goal in 63.77% of trials and recorded human collisions in 21.67%. NavThinker recorded a 39.09% human-collision rate in the same benchmark. The authors also report a 54.15% success rate and 23.34% collision rate after transferring the method to a different simulated environment, Social-MP3D, without retraining.
These are simulation results, not tests with a physical robot. The paper's failure analysis shows the limitation: when a pedestrian repeatedly blocks a narrow passage, Social-WM stays clear but does not decide to seek a detour, and the run eventually times out.
Sources
- Social-WM paper and revision record (arXiv v2)
- Full text of the Social-WM paper (arXiv HTML)
- Image: Daniel Lu, “Corridorbot door,” Wikimedia Commons, CC BY-SA 3.0. Illustrative corridor diagram, not an image from the study.
