Design and Control of a Bipedal Robotic Character

Ruben Grandia, Espen Knoop, Michael A. Hopkins, Georg Wiedebach, Jared Bishop, Steven Pickles, David Müller, Moritz Bächer

Seminal January 2025

TL;DR A 15.4 kg, 0.66 m bipedal robot character whose shape and motion come from animators, controlled by separate RL policies for standing, walking and short animated clips that an operator blends live from a two-joystick controller. Public deployments with up to three robots ran on the order of 10 hours without a fall.

What they do

Legged robots are usually designed for function. This one is designed for character: a small droid-like biped whose proportions, backward-bending knees and walking style were set by animators, with the mechanics and control built to serve that intent. The paper presents the whole pipeline, from animation tools and hardware to RL policies and a puppeteering interface for live shows.

How it works

Hardware. 15.4 kg, 0.66 m tall, 5 DoF per leg and a 4 DoF neck and head. Off-the-shelf actuators sit directly at the joints, joined by 3D-printed parts; there is no ankle roll actuator, so the foot soles are rounded. Antennas, LED eyes, a head lamp and speakers are open-loop “show functions”.
One policy per motion type. A perpetual policy stands while following head and torso commands, a periodic policy walks from velocity commands plus a phase signal, and episodic policies each imitate one animation clip. All are trained with PPO in Isaac Gym on imitation rewards against artist-authored references, with identified actuator models, domain randomization and random pushes.
Animation engine. At runtime it layers a looping background animation, operator-triggered clips and joystick offsets into policy commands, and switches policies. Walk-to-stand switches wait for the next double-support phase.
Puppeteering. The left joystick moves the torso while the head holds its gaze; the right joystick moves the gaze, with the torso following near the neck limits.

Results

Walking reaches 0.7 m/s forward, 0.4 m/s lateral and 1.8 rad/s turning, and tracks commanded velocities closely.
Mean absolute joint tracking error is 0.035 rad standing, 0.123 rad walking and 0.027-0.043 rad for the episodic motions (happy dance, excited motion, jump, tantrum). The jump reaches the knee torque limits.
Tracking velocity commands directly, without the animated reference, gives a foot-shuffling or stiff, upright gait.
Several public deployments with up to three robots gave on the order of 10 h of runtime without a single fall. The battery lasts at least 1 h.

Limitations

Each policy trains for about 2 days on an RTX 4090, so separate policies add overhead as episodic motions accumulate. The robot has no autonomy or perception: a puppeteer makes all high-level decisions, there is a limit to how many buttons an operator can use, and some audience feedback found the visible puppeteer distracting.

Figures

From Figure 1: Three instances of the robotic character performing an unscripted show. Apart from their theming they ... — From Figure 1: Three instances of the robotic character performing an unscripted show. Apart from their theming they are identical, and each is remote-controlled by a separate operator.

From Figure 2: The character design and control pipeline. Animation and mechatronic design iterate to define the char... — From Figure 2: The character design and control pipeline. Animation and mechatronic design iterate to define the character and its motions; reinforcement learning with imitation rewards turns them into command-conditioned control policies; at run-time the animation engine combines user input and animation content, drives the policies and synchronizes the show functions.

From Figure 3: Mechanical design. 5 degrees of freedom per leg and 4 in the neck and head. The torso holds the commun... — From Figure 3: Mechanical design. 5 degrees of freedom per leg and 4 in the neck and head. The torso holds the communication board, battery and IMU; the head holds the onboard PC, radio receiver and show functions (antennas, LED eyes, head lamp). Speakers sit in the head and the bottom of the torso.

From Figure 6: Joystick commands during standing. Posture control (left) moves the torso without affecting the gaze. ... — From Figure 6: Joystick commands during standing. Posture control (left) moves the torso without affecting the gaze. Gaze control (right) moves mainly the head, with additive torso movement to extend the range.

Abstract

Legged robots have achieved impressive feats in dynamic locomotion in challenging unstructured terrain. However, in entertainment applications, the design and control of these robots face additional challenges in appealing to human audiences. This work aims to unify expressive, artist-directed motions and robust dynamic mobility for legged robots. To this end, we introduce a new bipedal robot, designed with a focus on character-driven mechanical features. We present a reinforcement learning-based control architecture to robustly execute artistic motions conditioned on command signals. During runtime, these command signals are generated by an animation engine which composes and blends between multiple animation sources. Finally, an intuitive operator interface enables real-time show performances with the robot. The complete system results in a believable robotic character, and paves the way for enhanced human-robot engagement in various contexts, in entertainment robotics and beyond.

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