Contract v1.0Microduck runtime

The shared rhythm behind every move.

Behavior policies use one predictable interface: 61 observations in, 14 joint targets out, 50 times each second.

Keeping this shape fixed makes behaviors swappable. A policy can change from a walk to a recovery move without changing the daemon that runs it.

The contract below describes the values a policy receives and the targets it returns. Normalization statistics travel inside the ONNX graph.

01 / Observation vector

Eight groups make up the 61-dimensional input.

Index	Dims	Component	What it carries
0 – 13	14	Joint positions	Measured servo angles (rad) relative to the neutral pose.
14 – 27	14	Joint velocities	Measured angular velocity (rad/s) from encoder derivatives.
28 – 41	14	Previous actions	Targets sent on the previous control tick for temporal smoothness.
42 – 44	3	Projected gravity	Gravity unit vector in the trunk frame; upright is [0, 0, -1].
45 – 47	3	Angular velocity	Trunk gyroscope vector [wx, wy, wz] in rad/s.
48 – 50	3	Twist command	Operator intent: forward, lateral, and yaw velocity.
51 – 54	4	Head pose command	Neck and head orientation targets.
55 – 60	6	Body pose command	Torso translation and rotation offset [tx, ty, tz, roll, pitch, yaw].

Each output is a position offset around the home pose: target = home_pose + action_scale × action.

The beak is separate. Its 15th motor is driven through programmatic RPC or gamepad input, outside this fixed locomotion space.

Simulation uses the BAM M6 actuator model for the XL330 Dynamixel. It accounts for the details that matter on a small, light biped.

When several gestures are available, the daemon resolves them in this order:

roulade 〉 kick 〉 ground_pick 〉 sit/rise 〉 stand_by_magnitude 〉 walk