TLDR: AdaMimic is a new motion tracking algorithm that allows humanoid robots to adapt complex movements from a single reference motion. It achieves this by intelligently processing the motion into keyframes and using a two-stage reinforcement learning approach with “adapters” to adjust motion speed and actions. This enables robots to perform agile tasks like jumping and ball-hitting with high accuracy and adaptability in both simulations and on real hardware, overcoming limitations of previous methods that either sacrificed accuracy or required extensive data.
Humanoid robots are designed to mimic human movements and adapt them to various real-world situations. Imagine a robot learning to play tennis; it needs to reproduce a stroke pattern but adjust it for different ball positions. This ability, known as adaptable humanoid control, has been a significant challenge in robotics. Existing methods often face a trade-off: some offer good adaptability but sacrifice the accuracy of the original motion, while others achieve precise imitation but require a vast amount of training data or specific reference motions for every new scenario.
A new research paper, “Towards Adaptable Humanoid Control via Adaptive Motion Tracking,” introduces AdaMimic, a novel motion tracking algorithm that aims to combine the best of both worlds. Developed by Tao Huang, Huayi Wang, Junli Ren, Kangning Yin, Zirui Wang, Xiao Chen, Feiyu Jia, Wentao Zhang, Junfeng Long, Jingbo Wang, and Jiangmiao Pang from Shanghai AI Laboratory and Shanghai Jiao Tong University, AdaMimic enables humanoid robots to adapt from just a single reference motion while accurately preserving the intricate patterns of that motion. You can read the full paper here: Towards Adaptable Humanoid Control via Adaptive Motion Tracking.
How AdaMimic Works
The core innovation of AdaMimic lies in its ability to reduce data dependence while ensuring high adaptability. It achieves this through a two-stage training process and a clever way of preparing motion data:
- Augmented Dataset Creation: Instead of needing many examples, AdaMimic starts with a single reference motion. It then “sparsifies” this motion into keyframes – crucial moments in the movement (like the take-off or landing in a jump). These keyframes are lightly edited with minimal physical assumptions to create an augmented dataset. This process allows for variations in global movements (like how far a robot jumps) while strictly maintaining the local joint movements (the specific way a leg bends).
- Stage 1: Initial Motion Tracking: In the first stage, a policy is trained to track these sparse keyframes. This initial training uses a fixed speed for the motion. It employs a unique reward system: a “sparse global reward” that focuses on accurate alignment only at the keyframes, and a “dense local reward” that ensures the robot mimics the fine-grained joint movements throughout the motion. This dual reward system helps stabilize training and provides an initial capability to imitate the edited keyframe motions.
- Stage 2: Adapters for Flexible Adaptation: The real magic happens in the second stage. AdaMimic introduces two “adapters” that are jointly learned: a “phase adapter” and a “tracking adapter.” The phase adapter dynamically adjusts the motion’s speed, allowing for flexible time warping – essentially speeding up or slowing down parts of the movement as needed. The tracking adapter then refines the robot’s low-level actions based on these speed adjustments. This two-stage design enables the robot to adapt its motion timing and actions without altering the fundamental motion pattern, leading to improved accuracy and adaptability, especially in challenging scenarios.
Significant Improvements and Real-World Deployment
The researchers conducted extensive evaluations in both simulation and on the real-world Unitree G1 humanoid robot across various agile tasks, including different types of jumps and ball-hitting skills (badminton and tennis). AdaMimic consistently outperformed existing methods like AMP-Style, DeepMimic variants, and UniTracker.
In simulations, AdaMimic showed superior success rates, lower tracking errors (both global and local), and smoother motions, particularly in “hard adaptation” cases where significant adjustments were required. For instance, it could make the robot jump much farther or higher than the original motion, or adjust a tennis swing for a different ball position, all while maintaining the naturalness of the movement.
Crucially, AdaMimic’s trained policies were directly deployed on the Unitree G1 humanoid robot. Unlike some baseline methods that struggled with physically implausible motions or exhibited jerky, unstable movements on hardware, AdaMimic demonstrated robust and physically plausible execution across all tasks. This highlights the effectiveness and generality of the method in real-world applications, even under challenging adaptation conditions.
Also Read:
- SceneAdapt: Integrating Scene Understanding into Motion Generation
- Robots That Adapt: Adversarial Fine-Tuning for Robust Control
Future Directions
While AdaMimic represents a significant leap forward, the researchers acknowledge several areas for future improvement. These include extending the method to tasks without clear parametric forms, developing more general and automatic mechanisms for keyframe selection and editing, finding better strategies to utilize massive motion data, and incorporating perception to enable adaptive and responsive behaviors in interactive tasks like complex ball games.
AdaMimic offers a promising path towards more versatile and adaptable humanoid robots, capable of learning complex skills from minimal demonstrations and performing them robustly in diverse real-world environments.
