FLORE
A framework for learning force-reactive humanoid mobile manipulation that explicitly outputs a 6-dimensional force wrench as the policy output, combining a diffusion policy with a hybrid compliant whole-body controller to accomplish contact-rich tasks.
Timeline (last 90 days)
FLORE is a framework for learning force-reactive humanoid robot mobile manipulation, whose core idea is to make force part of the policy output rather than passively compensating for it afterward with a controller.
FLORE collects and outputs 6-dimensional wrenches, enabling the policy to express whole-body coordinated force intent.
FLORE uses two force-sensing UMI-style handheld grippers and wearable trackers for robot-free demonstration collection, recording hand motion, contact wrenches, and body motion.
FLORE's policy network takes vision, proprioception, and measured wrenches as input and outputs actions as well as force commands for each hand, using a diffusion policy to generate actions.
FLORE's execution layer employs a hybrid compliant controller with force tracking, using the integral of force error measured by real F/T sensors so that each hand precisely tracks force along the commanded direction while remaining compliant in other directions.
In simulation, FLORE's controller can track per-hand force commands of different magnitudes and directions while losing little whole-body motion accuracy.
In real-robot experiments, FLORE's force-reactive policy achieved an 80% success rate on three contact-intensive tasks, whereas a position policy trained on the same demonstration data and executed by the same controller achieved only a 26.7% success rate.
FLORE's current method still relies on specific handheld collection devices, and its cross-task generalization capability remains to be verified.
All Facts (8)
FLORE is a framework for learning force-reactive humanoid robot mobile manipulation, whose core idea is to make force part of the policy output rather than passively compensating for it afterward with a controller.
50%UnverifiedFLORE collects and outputs 6-dimensional wrenches, enabling the policy to express whole-body coordinated force intent.
50%UnverifiedFLORE uses two force-sensing UMI-style handheld grippers and wearable trackers for robot-free demonstration collection, recording hand motion, contact wrenches, and body motion.
50%UnverifiedFLORE's policy network takes vision, proprioception, and measured wrenches as input and outputs actions as well as force commands for each hand, using a diffusion policy to generate actions.
50%UnverifiedFLORE's execution layer employs a hybrid compliant controller with force tracking, using the integral of force error measured by real F/T sensors so that each hand precisely tracks force along the commanded direction while remaining compliant in other directions.
50%UnverifiedIn simulation, FLORE's controller can track per-hand force commands of different magnitudes and directions while losing little whole-body motion accuracy.
50%UnverifiedIn real-robot experiments, FLORE's force-reactive policy achieved an 80% success rate on three contact-intensive tasks, whereas a position policy trained on the same demonstration data and executed by the same controller achieved only a 26.7% success rate.
50%UnverifiedFLORE's current method still relies on specific handheld collection devices, and its cross-task generalization capability remains to be verified.
50%