Microduck BallKick 4096x6000

This repository stores the Microduck right foot kick policy reproduction of the every-embodied tutorial. The training is based on pollen-robotics/microduck_rl, and the submission is 29e887ecfbf5d37144759e5a9f8a176dfb83d547, with the task being Mjlab-BallKick-Flat-MicroDuck.

Training Records

Parallel environment: 4096
PPO iterations: 6000
Total transition count: 589,824,000
Training time: 2 hours 6 minutes 15 seconds
Final mean reward: 89.42
Final mean episode length: 250.0
Final ball_forward_velocity reward: 11.4758

These values only describe this run and should not be considered independent of code submission, task configuration, and random seeds as the cross-project ranking results.

Files

model_5999.pt: Final RSL-RL checkpoint.
ball_kick_right_6000.onnx: The right foot policy used for Microduck’s 61-dimensional observation and 14-dimensional action operation.
microduck_ballkick_visual_qa.json: Machine-readable records for training, browser closed loop, contact scanning, and demonstration videos.

Verification Value

textcf3f34ed68dec9137260d20f4e2cfdd507415df651d86fe31f2ac9daa8073c7c  model_5999.pt
bd03168c2554de6f95a69903143997c3ff85fec73c28a010225868216d3ea518  ball_kick_right_6000.onnx

Usage Boundaries

This policy does not receive image or ball position data, and only executes one blind kick. The tutorial webpage uses a separate visual detection interface and a finite state machine to identify the ball, approach, foot placement alignment, and policy switching. It provides a first-person view from the head, a second-person view at close range, and a third-person view centered. The detection boxes in the browser demo are obtained through MuJoCo truth values projection and are used to verify the control interface; they do not represent any training or evaluation of the visual detector.

For full descriptions, commands, and videos, see every-embodied’s Microduck reinforcement learning chapter.