Hip Exoskeleton: A controller trained in simulation with a human model with 304 muscles reduces the energy expenditure of walking by 18%, without individual adjustment

The MARL-Exo controller learned in front of a virtual body with various anatomys. Tested as it was in eight adults, it decreased their metabolic cost in all participants.

Adjusting an exoskeleton for each user, by measuring his breathing during long sessions, remains expensive. Yifei Yuan, Xianlian Zhou and their colleagues sought to obtain a single controller valid for all, trained entirely in simulation. Their MARL-Exo frame simultaneously teaches two agents: a complete human model with 304 muscle actuators, whose anatomy, approach and speed vary, and the policy of a bilateral hip exoskeleton, rewarded according to the metabolic energy spent by the muscles.

Confirmed results on real walkers

In simulation, the purely metabolic policy reduces muscle power by 13.8% compared to a walk without a device, compared to 8.8% when adding a constraint on mechanical power. Deployed without re-training on eight healthy adults walking at 1.25 m/s, it reduced the net metabolic cost by an average of 18.3 per cent compared to the passive apparatus in all participants.

Compared to walking without exoskeleton, the average gain drops to 7.4%, as raising the device itself costs about 13% of energy; However, six participants more than offset this additional cost, until walking with 24.2% less effort. The approach also allows testing of simulation assistance objectives before any human test.