microduck-running

This policy adds running to the Microduck at 1.6 m/s - 4x faster than the normal 0.4 m/s walking speed.

Command

The 13-D command block is [twist(3), head_pose(4), body_pose(6)].

slot	meaning
twist[0]	forward-velocity command; trained through 2.20 m/s, evaluated at 2.20 m/s
twist[1]	lateral command; use 0 (only a tiny ±0.02 m/s range was seen in training)
twist[2]	yaw-rate command; use 0 (only a tiny ±0.05 rad/s range was seen in training)
head_pose[0:4]	unused for this policy; zeros
body_pose[0:6]	unused for this policy; zeros

An exact-zero standing bucket made up 3% of training environments, so twist = [0, 0, 0] is the intended idle command. The speed objective rewards forward progress and can differ from the command; twist[0] is not a hard speed limiter. Do not begin a hardware test at the 2.20 m/s simulation command.

Try it in simulation

The training family is based on the private Vottivott/microduck_rl tag running-sim-v1, with the robustness continuation described below.

git clone https://github.com/Vottivott/microduck_rl.git
cd microduck_rl
git checkout running-sim-v1
uv sync

hf download HannesVonEssen/microduck-running policy.onnx --local-dir policies/running
uv run scripts/infer_policy.py \
  --walking policies/running/policy.onnx \
  --new-cmd-obs \
  --lin-vel-x 2.2

Contract

• input: obs, float32 [1, 61]
• output: actions, float32 [1, 14]
• control rate: 50 Hz
• action scale: 1.0, joint-position targets around MicroDuck HOME
• observation normalizer: baked into policy.onnx
• entry pose: standing
• robot: MicroDuck hardware revision 1, 14 Dynamixel XL330 servos
• policy type: perpetual flat-ground locomotion

Simulation evaluation

All results below use a 2.20 m/s command for 10 seconds. “Stress” adds random ±0.10 m/s planar velocity pushes every 3–6 seconds, trunk CoM offsets of ±10 mm, head CoM offsets of ±6 mm, and initial pitch/roll up to ±2 degrees.

model / condition	envs	survival	mean body-forward speed
plain, nominal	512	99.22%	1.6511 m/s
plain, stress	512	98.83%	1.6350 m/s
plain, stress + friction 0.7–1.8	512	96.68%	1.6424 m/s
backlash model, nominal	256	98.83%	1.6362 m/s
backlash model, stress	256	98.44%	1.6122 m/s

Heading control remains weak: nominal mean absolute heading error was 31.33° and final absolute heading error averaged 59.87° over the 10-second battery. The speed figures are simulation body-forward means, not real-hardware measurements.

Sim-to-real constraints

The policy retains deployment-oriented modeling from the walking recipe:

• BAM XL330 M6 voltage control, back-EMF, current limiting, battery variation from 6.5–8.2 V, and load-dependent voltage sag;
• 3–6 physics-step actuator delay, 0–1-step IMU delay, and one-control-step joint-velocity lag;
• foot friction randomized from 0.7–1.3 during training;
• trunk mass/inertia ±5%, armature ±10%, BAM joint friction ±10%, joint encoder bias ±0.015 rad, IMU misalignment up to 6°, and actor noise.

The fast-gait discovery run originally relaxed several constraints: flat terrain, no pushes, non-backlash dynamics, no wide CoM curriculum, weaker posture/motion/slip/action-rate regularization, and disabled pose tracking. The hardware continuation then progressively added:

• planar velocity pushes ±0.03 → ±0.06 → ±0.10 m/s;
• trunk CoM offsets ±3 → ±5 → ±8 mm;
• head CoM offsets ±3 → ±5 → ±6 mm;
• initial pitch/roll up to ±1 → ±1.5 → ±2°.

Backlash and the held-out ±10 mm trunk-CoM / high-grip combination were used for evaluation, not silently folded into nominal training. Remaining concerns include real impact loading, gearbox compliance, foot contact, heading drift, and sustained actuator current/temperature.