GaitSpan enables humanoid robots to run without relearning locomotion from scratch
Single policy that continuously transitions from walk to jog to run across terrains.
Humanoid robots that can walk typically have to relearn locomotion from scratch to jog or run, a significant inefficiency. GaitSpan, developed by Kwan-Yee Lin and colleagues, tackles this by treating a basic walking policy as a 'seed skill' that can be expanded into faster gaits. The framework uses three key mechanisms: rhythm generation that modulates the frozen walking policy with multiple internal clocks for new cadences; stride shaping that rewards dynamic patterns based on spring-loaded inverted pendulum physics; and residual adaptation that captures fine motion details. This approach avoids the need for separate expert policies or motion capture data.
In testing, GaitSpan achieved a single policy that covers walking, jogging, and running-like regimes across a continuous speed range, a first for humanoid robots. The policy transfers across different robot morphologies and works zero-shot on unseen simulated and real-world terrains. Compared to baselines using multi-experts or imitation from humans, GaitSpan learns faster and produces stronger gait performance. This work represents a significant step toward more versatile and energy-efficient humanoid locomotion, potentially enabling robots to dynamically adjust their gait in real-world applications like search and rescue, delivery, or industrial inspection.
- GaitSpan expands a single pretrained walking policy into jogging and running using three components: rhythm generation, stride shaping, and residual adaptation.
- It is the first humanoid locomotion system to deliver a single command-conditioned policy covering a continuous speed range from walk to run.
- The policy transfers zero-shot across morphologies and terrains, and learns faster than multi-expert or imitation-based baselines.
Why It Matters
Enables humanoid robots to dynamically adjust gait without retraining, crucial for real-world versatility and efficiency.