CrossLoco: Human Motion Driven Control of Legged Robots via Guided Unsupervised Reinforcement Learning

Tianyu Li, Hyunyoung Jung, Matthew C. Gombolay, Yong Kwon Cho, Sehoon Ha

Abstract

CrossLoco is a novel reinforcement learning framework designed to bridge the morphological gap between human motion and legged robot control. Unlike traditional supervised methods that require labor-intensive paired datasets, CrossLoco utilizes a guided unsupervised approach. By implementing a cycle-consistency reward system, the framework enables robots—such as quadrupeds—to discover and mimic the essence of human movements like running, dancing, and turning. Experimental results demonstrate that CrossLoco produces more diverse and agile motions compared to existing baselines, offering a scalable solution for intuitive robot teleoperation.

Introduction

Human Motion Driven Control (HMDC) represents a significant milestone in robotics, aiming to allow robots to follow human movements intuitively. However, a primary challenge lies in "morphological mismatch": a two-legged human and a four-legged robot have fundamentally different body structures. Establishing a direct correspondence between their poses is inherently ambiguous.

Existing solutions typically rely on supervised learning with massive, hand-labeled datasets or unsupervised methods that often struggle with the high-dimensional search space of complex robots. CrossLoco addresses these limitations by introducing a framework that discovers robot skills and establishes a semantic bridge between human kinematics and robot dynamics without requiring pre-labeled pairs.

Methodology: Guided Unsupervised Reinforcement Learning

The core innovation of CrossLoco is its Cycle-Consistency Engine, which treats reconstructibility as a proxy for semantic meaning.

The Cycle-Consistency Engine

The system trains two mapping networks alongside the robot's control policy:

  1. Robot2Human-Mapper (R2H): Predicts the equivalent human pose from a given robot pose.
  2. Human2Robot-Mapper (H2R): Maps the predicted human pose back to the original robot pose.

The reinforcement learning agent is rewarded when the cycle from robot to human and back to robot is consistent. This maximizes the mutual information between the two domains, ensuring that the robot's movements capture the "essence" of the human motion.

The CrossLoco architecture featuring the Cycle-Consistency Engine with R2H and H2R mappers.

Guided Reinforcement Learning

To prevent the robot from adopting overly conservative behaviors (like standing still to avoid penalties), the authors introduce a root-tracking reward. This provides a "hint" by encouraging the robot's main body (the root) to follow the general trajectory of the human. This guidance narrows the search space, allowing the cycle-consistency reward to focus on learning fine-grained joint movements and emergent gaits.

Experimental Setup

The framework was evaluated using the LaFAN1 dataset for human motion capture, focusing on 50 trajectories including walking, running, hopping, and dancing. The target platform was the Unitree Aliengo quadruped, simulated in the NVIDIA Isaac Gym physics engine. For extended functionality, the researchers integrated the Human Motion Diffusion Model (MDM) for text-to-motion tasks and Learned Motion Matching (LMM) for interactive joystick control.

Results and Discussion

CrossLoco was compared against DeepMimic, a manually-engineered baseline, and several ablation models.

Qualitative and Quantitative Performance

CrossLoco achieved superior scores in both Averaged Correspondence Reward (ACR) and Diversity (DIV). Qualitatively, the robot discovered distinct strategies for different motions, naturally switching between "trotting" and "galloping" gaits to match human styles.

Qualitative comparison of CrossLoco against DeepMimic and Tasked-RL across various motions.

In a user study with 15 subjects, participants preferred CrossLoco's motions 43% of the time over all other methods. Users were able to correctly match the robot's motion to the source human motion in 65.5% of cases, demonstrating the effectiveness of the semantic mapping.

Cross-Application Utility

The framework's versatility was showcased through Language2Robot applications, where text prompts like "strides swiftly" successfully triggered running behaviors. Additionally, the system supported real-time interactive control, allowing the robot to adjust its movement style dynamically via joystick input.

Limitations and Future Work

While successful, CrossLoco faces challenges with high-frequency motions like hopping, where the physical differences in ground impact management between humans and quadrupeds are most pronounced. Some extreme human maneuvers, such as sharp 180-degree turns, also push the physical limits of the quadrupedal morphology.

Future research will focus on extending the framework to long-horizon tasks and legged manipulation, such as a quadruped using a robotic arm to mimic human tool-reaching. The team also plans to validate the system on physical hardware in real-world environments.

Summary updated

BibTeX
@inproceedings{li2024crossloco,
  author = {Tianyu Li and Hyunyoung Jung and Matthew C. Gombolay and Yong Kwon Cho and Sehoon Ha},
  title = {{CrossLoco: Human Motion Driven Control of Legged Robots via Guided Unsupervised Reinforcement Learning}},
  booktitle = {International Conference on Learning Representations},
  year = {2024},
  url = {https://openreview.net/forum?id=i0FTEXdPbF},
}
Figure