GenPHRI
Physical human-robot interaction is hard to author by hand. GenPHRI turns natural-language task descriptions into complete simulated scenarios: a human, a setting, a robot, and the motion that connects them.
Select a scenario to watch its motion, rotate the scene, and inspect its task specification.
Prompt:
Generated motion can be deployed directly for a rapid physical preview, or used to train a vision policy for closed-loop interaction. Both routes start with the same scenario.
Fit a human body mesh from a single image, then evaluate the generated motion on that body. No demonstration collection or policy training.
A fast way to inspect the intended behavior. Contact coverage depends on the accuracy of the fitted body mesh.
4x speed
Collect demonstrations across diverse simulated bodies and train a visuomotor policy that responds to observations throughout the interaction.
Higher contact coverage through closed-loop control. Trained entirely in simulation, with no real-world fine-tuning.
4x speed
In a 12-participant study, learned policies achieved 80% bathing and 79% scratching target completion without instructed arm movement.
Direct execution and policy learning had no statistically significant difference in bathing prompt-adherence ratings, while learned policies achieved substantially higher contact coverage.
Bathing used 15 evenly distributed removable targets across the back and legs; scratching used five contact markers along the forearm. Participants rated adherence to the original task prompt on a seven-point scale.
Each policy was trained on 1,000 simulated demonstrations across 100 human bodies. Simulated evaluation used 20 held-out bodies and two random seeds per body. Mean simulated completion was 86% for bathing and 92% for scratching. Simulated targets used a 2 cm proximity criterion, so the simulation and physical measurements are not identical.
Scratching with instructed arm movement reached 63% completion in real-world trials, highlighting a remaining challenge in adapting to human motion.
Task: Scalp Scratching
Loading scene…
Loading scene…Attempts are exhausted. Accept the earlier placement that passes reachability and clearance checks, despite its less-preferred approach from behind.
Task: Head and Neck Towel Drying
Loading scene…
Loading scene…The hands remain beside the hips across repeated attempts. Retry with explicit guidance to bend the elbows forward and bring the hands onto the lap.
Task: Right Foot Sponge Bathing
Loading scene…
Loading scene…The last candidate improved, but reach remains limited. Continue refinement with guidance to move closer to the foot.
Task: Mid-Back Percussion
Loading scene…
Loading scene…The current human placement blocks access. Revisit it before trying robot placement again.
Task prompt: A person lies face-down on a bed, arms resting alongside the torso and legs extended straight. A robot bathes them with a sponge, wiping the whole back in several passes and then the back of each leg down to the ankle, and it pulls the sponge away from the body when it finishes. The robot must not touch the buttocks or the area between the legs.