Humanoids and exoskeletons: balance, whole-body control, reality

Every exoskeleton user is different

What happens. An exoskeleton must help the wearer move, not fight them. But body size, strength, gait and fatigue differ from person to person and change during the day.

What it costs. A controller tuned for one person feels wrong on the next. Wearers push against the device, comfort drops, and acceptance suffers. Fitting takes a therapist or engineer for every user.

Why the usual fix fails. Fixed presets for a few body types are a start but miss the individual. Manual retuning after every session is not practical in daily use.

How a fixed-price remote diagnosis checks it. From recorded sessions (joint angles, torques, interaction forces), I check how much of the difference between users can be learned and adapted. The fixed-price remote diagnosis tells you whether an adaptive, model-based controller could fit itself to the wearer, and what sensing it would need.

Legged and wheeled-legged robots fall on uneven terrain

What happens. Robots with legs or wheels on legs have to keep their balance on slopes, steps and slippery floors, while carrying loads that shift. They are pushed to their limits by the contact with the ground.

What it costs. Falls damage the robot and stop the work. Developers make the gait slow and cautious, and the robot does not do what it was bought to do.

Why the usual fix fails. Tuning the gait for one terrain in the lab does not carry over to others. Reinforcement learning trained in simulation can fail on real ground in ways nobody can explain.

How a fixed-price remote diagnosis checks it. From test logs (body motion, contact forces, foot slip), I check what the controller does at the limit of balance. The fixed-price remote diagnosis tells you whether predictive control with explicit limits on contact forces and joint torques would hold the robot where a tuned gait fails.

Whole-body control on embedded hardware is hard to staff

What happens. A humanoid has dozens of joints that must be coordinated while it balances. The controller has to run fast, on computing hardware inside the robot that is far smaller than a workstation.

What it costs. The skills to combine whole-body predictive control with real-time code are rare. Roles stay open, and the team falls back on a simpler controller with less capability.

Why the usual fix fails. Running the full controller off-board is not an option for a robot that must walk free. A neural policy trained in simulation is hard to certify and to debug.

How a fixed-price remote diagnosis checks it. I review your controller formulation and its measured computing time on the target. The fixed-price remote diagnosis tells you where a reduced model or faster solver would bring the controller into the time budget, and how to test it in steps.

The gap between simulation and real hardware

What happens. A robot is developed in simulation, where the controller works well. On the real hardware, with friction, backlash, flexible parts and sensor delay, it behaves differently.

What it costs. The team loses weeks of hardware time tuning what the simulation promised. Tests on real hardware are risky, because a bad controller can break the robot.

Why the usual fix fails. Making the simulation more detailed does not close the gap if the parameters are guesses. Tuning gains on the robot by hand is slow and is lost at the next hardware change.

How a fixed-price remote diagnosis checks it. From recorded motions of the real robot, I check how well a model with identified parameters reproduces them. The fixed-price remote diagnosis tells you whether system identification and state estimation would close most of the gap, and what motions to record first.

Locomotion with trajectory optimisation and predictive control is hard to staff

What happens. Walking needs the robot to plan its steps ahead and adjust them as the ground and the body move. This is done with trajectory optimisation and predictive control, a field with few experienced people.

What it costs. Teams find it hard to hire for it. Without it, the robot walks slowly, and each new gait is a research project.

Why the usual fix fails. Gaits recorded from human motion and replayed do not respond to a push or an unexpected step. Reinforcement learning alone cannot guarantee that limits on force and torque are kept.

How a fixed-price remote diagnosis checks it. I review how your walking controller plans and replans, with its logs. The fixed-price remote diagnosis tells you which gaits a predictive formulation could cover, and what the team would have to build around it.

Series production of humanoids needs control and testing at scale

What happens. Several companies plan series production of humanoid robots within a few years. A prototype that walks in the lab has to become a product that thousands of units walk the same way.

What it costs. Balance, whole-body control and the testing of every unit before it leaves the factory become a production problem, not a research one. Unit-to-unit differences in joints and drives show up in the gait.

Why the usual fix fails. Hand tuning each robot does not scale. A test that takes an hour per unit is too slow for a line.

How a fixed-price remote diagnosis checks it. From the data of a few prototype units, I check how the variation between units affects the controller and which parts can be estimated and corrected in the factory. The fixed-price remote diagnosis tells you what a production test and calibration of control would need to look like.