The MIT Humanoid Robot: Design, Motion Planning, and Control For Acrobatic Behaviors

Matthew Chignoli, Donghyun Kim, Elijah Stanger-Jones, Sangbae Kim

Seminal April 2021

TL;DR A design for a 0.7 m, 21 kg humanoid built on MIT Cheetah-style proprioceptive actuators, with a kino-dynamic planner that respects actuator limits and an MPC plus whole-body landing controller. In a simulation that models the measured actuators and battery, it does back flips, front flips and spinning jumps.

What they do

The paper presents the MIT Humanoid, a small humanoid designed for acrobatics, together with the planner and controller meant to drive it. The robot applies the MIT Cheetah design principles (torque-dense electric motors, high-bandwidth force control, backdrivability) to a humanoid for the first time. Flips, spins and jumps are shown in a dynamics simulation that includes the measured actuator and battery behavior; the full robot was not yet assembled.

How it works

Hardware. About 0.7 m tall and 21 kg, with 5-DoF legs and 3-DoF arms; 75.6% of the mass sits in the torso, shoulders and hips. Two new motor modules, U10 (619 g) and U12 (1174 g), derive from the Mini Cheetah actuator. There are no torque or force sensors.
Actuator model. A custom dynamometer measures torque constants and peak torque. The model adds back-EMF limits and battery voltage droop, so torque limits are recomputed at every simulation step.
Actuation-aware kino-dynamic planner (AAKD). Centroidal-dynamics trajectory optimization for takeoff, with an added constraint that maps ground reaction forces to joint torques through a partly linearized contact Jacobian and bounds them by a linear torque-speed limit.
Landing. A lumped-mass MPC (10 Hz, 1.5 s horizon) supplies reaction forces and a body orientation reference to whole-body impulse control running at 500 Hz, which ranks body orientation above centroidal momentum.

Results

On the dynamometer the U10 saturates at about 31 Nm; the U12 reaches its 68 Nm gearbox limit before saturating. With belt gearing the knee is rated at 136 Nm.
In a 180 degree spinning jump, AAKD keeps every joint within its limit while standard kino-dynamic planning does not. Solve times are 3.10-8.77 s against 2.43-6.51 s, 28.2% longer on average.
Back flip off a 0.4 m platform, standing front flip, spin and forward and lateral jumps all run in simulation, with torques that virtually never exceed the measured actuator limits. Stabilizing arm swings emerge from the centroidal momentum task, not the planner.

Limitations

All acrobatic results are simulation only. Transfer to hardware rests on three stated assumptions: accurate inertial properties, an adequate rigid ground-contact model and adequate state estimation. The motion selector simply interpolates to a specified terminal state, and the landing MPC assumes near-constant body orientation and ignores precession and nutation.

Figures

From Figure 1: MIT Humanoid backflip. The proposed system performs a back flip off a 0.4 m platform in simulation.

From Figure 2: Configuration of the MIT Humanoid robot. MIT humanoid has 5-DoF legs and 3-DoF arms. The ankle joint h... — From Figure 2: Configuration of the MIT Humanoid robot. MIT humanoid has 5-DoF legs and 3-DoF arms. The ankle joint has only pitch directional actuation. Two contact sensors are located at toe and heel of each foot.

From Figure 5: Planning and Control Framework. Kino-dynamic planning, MPC-based landing control, and WBIC framework t... — From Figure 5: Planning and Control Framework. Kino-dynamic planning, MPC-based landing control, and WBIC framework to achieve dynamic aerial motions of a humanoid robot.

From Figure 7: Acrobatic Humanoid. MIT Humanoid performing a 180 degree spinning jump (top) and a standing front flip... — From Figure 7: Acrobatic Humanoid. MIT Humanoid performing a 180 degree spinning jump (top) and a standing front flip (bottom).

Abstract

Demonstrating acrobatic behavior of a humanoid robot such as flips and spinning jumps requires systematic approaches across hardware design, motion planning, and control. In this paper, we present a new humanoid robot design, an actuator-aware kino-dynamic motion planner, and a landing controller as part of a practical system design for highly dynamic motion control of the humanoid robot. To achieve the impulsive motions, we develop two new proprioceptive actuators and experimentally evaluate their performance using our custom-designed dynamometer. The actuator's torque, velocity, and power limits are reflected in our kino-dynamic motion planner by approximating the configuration-dependent reaction force limits and in our dynamics simulator by including actuator dynamics along with the robot's full-body dynamics. For the landing control, we effectively integrate model-predictive control and whole-body impulse control by connecting them in a dynamically consistent way to accomplish both the long-time horizon optimal control and high-bandwidth full-body dynamics-based feedback. Actuators' torque output over the entire motion are validated based on the velocity-torque model including battery voltage droop and back-EMF voltage. With the carefully designed hardware and control framework, we successfully demonstrate dynamic behaviors such as back flips, front flips, and spinning jumps in our realistic dynamics simulation.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong