FLoRe gives a humanoid a force command alongside its movements

FLoRe, a preprint posted on 26 October, completed 24 of 30 physical trials involving gantry pushing, box rotation and board wiping. A matched position-only policy completed eight. These were laboratory tests, with ten trials and a separate policy per task.[1]

How a demonstration becomes an action

Two earlier ideas help explain the design. The Universal Manipulation Interface, presented in 2024, uses handheld grippers to collect human demonstrations for training robots. The demonstrator can work without a robot at the collection site. Its learning interface represents movements relative to the current pose and accounts for delays between observations and actions, helping bridge the gap between a person demonstrating a task and a machine executing it. The researchers demonstrated policies transferable across different robot platforms, using the recorded interaction to teach behaviour.[2]

Diffusion Policy, introduced in 2023, supplies another ingredient: a way to generate robot actions by progressively refining noise into a proposed action sequence, conditioned on observations. It can represent several plausible ways to complete a movement, and its receding-horizon design repeatedly updates the plan as new observations arrive. This matters for manipulation because a useful command is a coordinated sequence involving several joints, rather than an isolated choice of the next coordinate. The original work combined visual conditioning with this sequence-based approach.[3]

FLoRe records contact forces during handheld demonstrations. Its diffusion policy senses contact loads and predicts hand forces alongside movement. A whole-body controller adjusts hand targets using measured force errors while maintaining balance.[1]

The revealing experiment: disconnect the command

The interpretation is practical: teaching a model to predict a useful quantity is only part of the design problem. The next component must be able to use that prediction. Otherwise, adding a new sensor or training target can improve the model’s internal description without giving the machine a corresponding action.

This is also a useful distinction when judging robot demonstrations. A smooth trajectory shows where a machine travelled. Successful manipulation additionally depends on what happened at the interface with the object. A system designer needs to decide which properties of that interaction should be measured, represented and commanded, rather than assuming that more accurate movement will settle every contact problem.

For FLoRe, withholding the force command produced a larger drop than removing force observations. The result points toward an action interface that carries the intended contact through to execution.[1]

Report an error or suggest a correction