What can you try?
Five ducks in one scene, each walking independently with the pretrained BEST_WALK_ONNX and bumping into the others. Real MuJoCo physics, no scripted motion.
Ideas to try on your own Mac, from easy to hard. Times are rough estimates at our M5’s speed (≈5,600 steps/s). Anything not marked “we tried” we have not tried either; we don’t know the result.
Easy (minutes to 1 hour)
Section titled “Easy (minutes to 1 hour)”1. Break the world, measure the policy · 5 min · we tried
Section titled “1. Break the world, measure the policy · 5 min · we tried”No training, measurement only. Change torque, friction or mass with eval_world.py. Question: which change is the duck most sensitive to? What happens at mass ×1.5? At friction 0.2?
2. Walk at a different speed · 5 min
Section titled “2. Walk at a different speed · 5 min”eval_checkpoint.py --vx 0.3 (twice the command) or --vx 0 (stand). Commands are sampled randomly in training, so the policy should know several speeds. Does distance grow with the command?
3. Write your own reward · 20–40 min · we tried
Section titled “3. Write your own reward · 20–40 min · we tried”Add a line to VARIANTS as in part 2. Ideas: “be quiet” (action_rate −5), “forget imitation” (imitation 0, rest official), “walk backwards” (fwd_vel −10). Guess which hack the duck will find, then look.
4. Another seed · 30 min (10 M) – 3.6 h (72 M)
Section titled “4. Another seed · 30 min (10 M) – 3.6 h (72 M)”train_mac.py --seed 1. Our duck lifted its feet at 68.4 M; when does yours? This shows why “one run is not statistics” matters.
Medium (a few hours)
Section titled “Medium (a few hours)”5. Train to the end of the official recipe · ~7.5 h
Section titled “5. Train to the end of the official recipe · ~7.5 h”--num_timesteps 150000000. We stopped at 48 %. Does it approach BEST_WALK_ONNX’s 58.5 cm at 150 M? An overnight run.
6. Train with a weak motor · 1–2 h
Section titled “6. Train with a weak motor · 1–2 h”The experiment of part 3: lower the torque limit to the catalog rating, or randomize it, with scripts/sim2real/train_sim2real.py --force 1.86 or --force_rand 1.40 1.90. For us 72 M steps were not enough; it walked at 88.8 M and forgot after 170 M. Try another seed or a fixed 1.86, and watch the checkpoints with watch_ckpts.py.
7. Rough terrain · a few hours
Section titled “7. Rough terrain · a few hours”Open_Duck_Playground has a rough_terrain_backlash task (at the pinned commit only the backlash version exists). You need to change the joystick.Joystick(task="flat_terrain") line in train_mac.py. What does a flat-floor duck do on rough ground, and vice versa?
Hard (a project)
Section titled “Hard (a project)”8. Teach it to get up
Section titled “8. Teach it to get up”Hugging Face’s Microduck can get up after a fall; the Open Duck environment has no get-up task. You need episodes that start on the floor, a trunk-height reward and a trunk that collides with the floor (the collision boxes from part 2 are ready). soyeon24’s open-loop attempt reached a crouch (10.6 cm of a 15 cm stance). Getting up is harder than walking.
9. Teach a show move (Disney style)
Section titled “9. Teach a show move (Disney style)”Besides walking, Disney’s BDX robot learns short moves like “happy dance”, “jump” or “tantrum” with an imitation reward (arXiv 2501.05204; ~22 days on one RTX 4090). The small version: make a reference motion from your own keyframes (Open_Duck_reference_motion_generator) and tie the imitation reward to it.
10. Move it to a real robot
Section titled “10. Move it to a real robot”Our ONNX files have the upstream interface (obs 1×101 → continuous_actions), so they should run with Open_Duck_Mini_Runtime. We have not tried hardware. Do experiment 6 first: a policy trained at 3.23 N·m in simulation covers half the distance at the 1.86 N·m catalog rating; we don’t know which one the real servo is closer to.