Proto (2)

Proto (2) goes after 1:1 human joint parity. Where Proto (1) proved the sim to real pipeline on cheap hardware, Proto (2) replaces the hardware underneath it: 28 degrees of freedom, real actuators, and a control stack that can hold a pose against forces Proto (1) has no way to measure.

Degrees of freedom

28 in total. Everything Proto (1) has, plus the joints it was missing:

• Elbows and wrists.
• Hip abduction, so it can finally step sideways and recover laterally.
• Torso yaw and pitch.
• Tendon driven five finger hands.

Actuation

• Field oriented control on quasi-direct-drive BLDC motors, replacing the serial bus servos entirely. The reason is torque density: Proto (1) lives inside a 0.50 N·m sustained budget, and that budget is what caps everything it can do.
• 48 V CAN bus.
• Per joint current sense torque estimation, so the robot knows what every joint is actually pushing rather than inferring it.
• Series elastic force feedback.

Control

• 1 kHz RTOS control loop under a Jetson Orin, twenty times the rate Proto (1) runs at.
• Whole body reinforcement learning on the same sim to real pipeline that was proven on the low cost build: a residual policy over a model based gait, domain randomized.
• Dynamic balancing across different mass distributions. Proto (2) should stay upright when it is carrying something, when the load is off centre, and when the mass moves while it walks. Proto (1) has no mechanism for any of that.

Sensing

This is the biggest change. Proto (1) runs an external IMU with no filtering at all. Proto (2) gets a real sensing stack: filtered inertial measurement, per joint torque from current sense, series elastic force feedback at the limbs, and the state estimation needed to close a balance loop at 1 kHz instead of riding a scripted gait phase.

Power

Custom battery management with cell balancing, undervoltage lockout, and cycle logging. Proto (1) runs a 4S LiPo behind a manual toggle and trusts the operator.