Yifei Yao
ARIEL: An Agent-Ready Humanoid Behavior Model with In-Context Execution Memory
ARIEL gives an agent a simple interface to whole-body control. Choose a demonstration, load it into memory, then recall the motion or steer its style with velocity and heading commands. The controller keeps executing between agent decisions.
Demos
Three motions, replayed from memory
Keep a style while changing direction
Why agent-ready?
- Choose the behavior. A demonstration becomes reusable execution memory. The agent can load it while the robot stands, then recall it later.
- Steer with simple commands. Once a style is loaded, the agent steers with velocity and heading commands, without generating a full-body trajectory at every step.
- Maintain continuous control. The controller uses retained memory, robot feedback and the latest command to continue low-level control while the agent prepares its next update.
These interfaces connect intermittent agent decisions to continuous motor control. Robustness to delays and abrupt command changes is still being evaluated.
ARIEL builds on GBC with a 4.24M-parameter controller. Its shared Transformer and fixed-size memory provide a path to more capacity and motion data; scaling experiments are ongoing.
Deployment measurements
An earlier streaming prototype measured 1.09 ms median inference with 4.21M weights, versus 13.92 ms and 49.73M weights for SONIC v1.1: about 12.7× lower latency and 11.8× fewer weights on the same Xeon 8358P CPU, using single-thread ONNX Runtime, 300 warm-up calls and 3,000 timed calls. These are historical measurements; deployment timing for the current memory controller is being evaluated separately.