Abstract
Developing versatile humanoid robots capable of tracking diverse, highly dynamic motions remains a significant challenge due to data inefficiencies and the inherent difficulty of balancing stability with agility. We present EGM (Efficient General Mimic), a framework that enables a humanoid robot to learn a robust, unified policy for a vast library of motions—including spinning jumps and sudden stops—using only 4.08 hours of curated training data. EGM achieves this through a novel bin-based sampling strategy and a body-part-decoupled Mixture-of-Experts (MoE) architecture, demonstrating superior generalization across nearly 50 hours of unseen test motions.
Introduction
Humanoid robots are designed for versatility, yet teaching them to switch seamlessly between disparate motions like running, throwing, and dancing is complex. Current research often relies on massive datasets (e.g., AMASS) that are frequently redundant or contain low-quality retargeting, which can hinder the learning of complex maneuvers. Furthermore, standard neural architectures often struggle with the "stability-agility trade-off," where the lower body's control requirements for balance conflict with the upper body's requirements for specific task execution. EGM addresses these issues by focusing on high-quality data curation and specialized architectural decoupling.
The EGM Framework
The EGM framework introduces three core innovations to optimize learning efficiency and tracking performance:
1. Bin-based Cross-motion Curriculum Adaptive Sampling (BCCAS)
Rather than sampling entire motion clips, EGM segments motions into 1-second "bins." The system utilizes an Exponential Moving Average (EMA) to monitor the tracking error of each bin. A curriculum-based strategy then shifts sampling probability toward high-error (difficult) bins as training progresses, ensuring the model spends more time mastering complex maneuvers like climbing or sudden transitions.
2. Composite Decoupled Mixture-of-Experts (CDMoE)
As illustrated in the framework overview, the CDMoE architecture splits experts into distinct groups for the upper and lower body. Within these groups, the model further decouples Shared Experts, which handle fundamental features like balance, from Orthogonal Experts, which specialize in motion-specific nuances. This separation ensures that learning a complex new skill does not degrade the robot's foundational stability.
3. Three-Stage Curriculum Training
The policy is developed through a tiered approach:
- Stage 1: Basic RL establishes fundamental tracking in simple environments.
- Stage 2: Complex RL introduces domain randomization and uneven terrains to build robustness.
- Stage 3: Policy Distillation uses DAgger to convert the privileged teacher policy into a student policy deployable with realistic sensor data.
Data Curation Strategy
A key insight of this work is that data quality outweighs quantity. The researchers compared a manually screened, high-quality subset ($D_{small}$, 4.08 hours) against a larger, automatically filtered set ($D_{large}$, 27.47 hours). Surprisingly, the policy trained on the smaller, high-fidelity set outperformed the larger one, suggesting that a diverse "gold" set is more valuable for general motion tracking than a massive, noisy dataset.
Experimental Results
EGM significantly outperforms the state-of-the-art GMT baseline across all primary metrics, including tracking error ($E_{mpkpe}$) and velocity distribution ($E_{vel\_dist}$).
Key Findings:
- Robustness and Stability: In high-dynamic tests, such as transitioning from a full run to a sudden stop, EGM maintained perfect stability where other MoE variants either became overly conservative or failed entirely.
- Generalization: Despite the limited training data, EGM generalized to over 49 hours of unseen motions, maintaining low tracking errors even for sequences exceeding 50 seconds in duration.
- Efficiency: The framework achieved these results using approximately 1/7th the training data volume typically required by comparable humanoid control research.
Conclusion and Future Work
EGM demonstrates that by combining intelligent data sampling with a decoupled expert architecture, humanoid robots can learn highly dynamic and diverse motion libraries with remarkable efficiency. While the current results are focused on simulation, the inclusion of domain randomization in the training pipeline prepares the policy for future deployment on physical hardware, such as the Unitree G1. Future research will focus on narrowing the performance gap during policy distillation and automating the identification of high-quality training data.