Asimov 1 ping-pong balance
humanoid control
1 min readupdated
A reinforcement learning (RL, trial-and-error agent training method) policy trained in MuJoCo (physics simulation software environment) and transferred to the real Asimov 1 humanoid robot. The model analyzes video from the head camera and IMU (acceleration and rotation sensor) readings in real-time, controlling right-arm movements to keep a tennis ball on a racket. The development is designed to test dynamic control in robotics and bridge the sim-to-real gap.
Key features
- PPO (proximal policy optimization algorithm) training in MuJoCo and transfer to real hardware.
- Use of a Kalman filter-based predictor to compensate for camera latency.
- Support for autonomous robot standing without a gantry frame.
- Right-arm control at 50 Hz processing joint and gyro observations.
What you can do
- Run the control pipeline simulation in MuJoCo using the repository scripts.
- Test model performance evaluation in an environment close to the real robot.
- Explore training configurations, logs, and checkpoints of the albert_v3_final model.
Example
from asimov_pp.ppo import load_policy
policy = load_policy("outputs/albert_v3_final/best.pt")
action = policy(obs)Read more
DetailsBenchmarks: up to 57 seconds of continuous balancing in free mode.
Requirements: Python, PyTorch, MuJoCo, access to the Asimov 1 robot SDK.
Features: training with simulated camera latency and real robot sensor noise.










Comments