Berkeley Humanoid: A Research Platform for Learning-based Control

Qiayuan Liao, Bike Zhang, Xuanyu Huang, Xiaoyu Huang, Zhongyu Li, Koushil Sreenath

★ Must read July 2024

The robot walking outdoors on the UC Berkeley campus.

TL;DR A 16 kg, 0.85 m humanoid built in-house for about USD 10,000 and designed so that simulation matches the hardware. A plain MLP policy trained with light domain randomization transfers zero-shot, walking 364 m across campus and hopping on one leg.

What they do

Berkeley Humanoid is a mid-scale humanoid research platform built almost entirely in-house: actuators, motor drivers, IMU, communication and power boards. It weighs 16 kg, stands 0.85 m tall, has two 6-DoF legs and costs about USD 10,000 without arms. The design goal is a narrow sim-to-real gap, so that a very simple RL controller is enough for robust outdoor locomotion.

How it works

Simulation-friendly hardware. No springs, dampers or closed kinematic chains. Custom quasi-direct-drive actuators (9:1 planetary gearbox, crossed-roller bearing) are mounted directly as joints; only the ankle pitch uses a linkage, with a linear joint-actuator mapping. EtherCAT keeps latency at 0.5-2 ms, so the actuators are simulated as delay-free torque sources, at more than 90,000 simulation steps per second on one GPU.
Reliable and cheap. Aluminum and steel structure, cables routed through hollow shafts, and a cell-phone-grade IMU that costs less than a dollar.
Minimal controller. An MLP actor (512, 256, 128) trained with PPO in Isaac Lab sees only the current state: no history, phase signal, reference motion or teacher-student training. It outputs joint position targets at 50 Hz.
Targeted randomization. Mechanical parameters come from CAD and per-actuator friction tests, so their ranges are small; only contact properties and pushes are randomized widely. Motor strength and PD gains are not randomized.

Results

Walks on grass, brick, asphalt, stairs and unpaved trails, and recovers from kicks. It covered 364 m on campus in 10 minutes and climbed a trail with a 20 degree average incline for 96 m and 10.5 m of elevation gain.
Velocity tracking error over 60 s: 0.051 m/s in simulation vs 0.058 m/s on hardware (sagittal), and 0.086 vs 0.1156 m/s (lateral).
A hopping policy with the same setup except for rewards hops on two legs and on one leg (with a mostly slack safety rope).
38 recorded falls caused no hardware damage beyond two failures from loose screws and glue; resets take 3 to 5 seconds.

Limitations

The robot has no arms yet. Range of motion, backlash, weight and mechanical strength need further hardware iterations. For more dynamic motion, the authors say the torque-current nonlinearity near saturation needs system identification, and motor operating region and heat protection should be simulated during training.

Videos

Single-legged hopping, with a mostly slack safety rope.

Double-legged hopping in the lab.

Figures

From Figure 1: CAD design, training in Isaac Lab and sim-to-real deployment of the custom-built humanoid with a learn... — From Figure 1: CAD design, training in Isaac Lab and sim-to-real deployment of the custom-built humanoid with a learning-based controller.

From Figure 2: Main components of the robot: computer, power management board, batteries and IMU in the torso, and th... — From Figure 2: Main components of the robot: computer, power management board, batteries and IMU in the torso, and the joints and actuators of the legs.

From Figure 3: Exploded view of one of the custom actuators, with motor driver, rotor, stator, planetary gear reducer... — From Figure 3: Exploded view of one of the custom actuators, with motor driver, rotor, stator, planetary gear reducer, crossed-roller bearing and hollow shaft.

From Figure 7: Recorded GPS track of a long-distance walk on the UC Berkeley campus.

Abstract

We introduce Berkeley Humanoid, a reliable and low-cost mid-scale humanoid research platform for learning-based control. Our lightweight, in-house-built robot is designed specifically for learning algorithms with low simulation complexity, anthropomorphic motion, and high reliability against falls. The robot's narrow sim-to-real gap enables agile and robust locomotion across various terrains in outdoor environments, achieved with a simple reinforcement learning controller using light domain randomization. Furthermore, we demonstrate the robot traversing for hundreds of meters, walking on a steep unpaved trail, and hopping with single and double legs as a testimony to its high performance in dynamical walking. Capable of omnidirectional locomotion and withstanding large perturbations with a compact setup, our system aims for scalable, sim-to-real deployment of learning-based humanoid systems. Please check http://berkeley-humanoid.com for more details.

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