News

Most ‘rehabotics’ (rehab-robotics) so far has been good at one thing and poor at another. It can move a limb along a set path well enough, but it really struggles with the part a human therapist would do instinctively; sensing how much resistance to give, when to yield, when to push and adjusting all of that moment to moment by feel, etc. A group at MIT, working with the Technical University of Munich, has built a system aimed at exactly that gap. The team (Johannes Lachner, now at Purdue, and Noah Geiger, with Neville Hogan’s Newman Laboratory at MIT) set a clear design principle; extend therapists, not replace them, and let a therapist train their own robot. Their paper, “Diffusion-Based Impedance Learning for Contact-Rich Manipulation Tasks,” is in IEEE Transactions on Robotics.

The method fuses two established ideas. The first is impedance control, a long-standing robotics approach (Hogan’s own field) that governs how a robot responds to physical contact by setting its stiffness and damping; how rigidly or softly it reacts to being pushed. It’s stable and safe, but conventionally an engineer has to hand-tune those stiffness and damping values for each task, which doesn’t scale to real, varied rehabilitation. The second is a generative diffusion model, the same class of AI used to generate images, but here it’s repurposed to generate behaviour rather than pixels.

What the diffusion model produces works like this: from the forces measured at the robot’s contact point (the external ‘wrenches’), a transformer-based diffusion model reconstructs what the researchers call a simulated Zero-Force Trajectory; the path the limb would follow if it were moving freely, with no resistance. An energy-based estimator then compares that intended path with what’s actually happening under contact, and adapts the robot’s stiffness and damping in real time, tightening along the direction of the task and softening along the axes that are irrelevant to it. So instead of a fixed, pre-tuned response, the assistance is continuously reshaped by the forces the patient is producing.

The performance figures show how well this works. Trained on only tens of thousands of samples, gathered by teleoperating the robot through both rehabilitation movements and an obstacle ‘parkour’ using an Apple Vision Pro headset, the model reached sub-millimetre positional and sub-degree rotational accuracy, ran in real-time torque control on a standard KUKA LBR iiwa arm, and generalised to tasks it had never seen; a 30/30 success rate inserting cylindrical, square and star-shaped pegs it wasn’t even trained on. In other words, it had learned something about physical interaction, not just memorised a set of movements – which is quite amazing.

In an ongoing study with Professor Cristina Piazza’s group at TUM and the Pfennigparade rehabilitation centre in Munich, physiotherapists and occupational therapists are treating patients while wearing force-sensing gloves and being filmed. That data is being used to build models that capture the physical style of an individual therapist, with the eventual aim of therapist-specific models, evaluated against the same patients who previously had manual therapy. So the robot learns from expert human handling rather than from an abstract idealisation of it,  then delivers high-volume, force-adaptive practice to stroke survivors under supervision… this seems to me to be to be very useful – and on the right track to make robotic additions actually helpful addition, rather than promising so much and actually not really delivering anything much more than a therapist can do very well a without vastly expensive robot…