Tagged: training
Every guide, note, and policy on YouDuck tagged「training」.
- Backlash, and which side the encoder sits on
Gear play. Microduck models ±1° per servo, and the point is that the encoder reads through it.
- The BAM actuator model
Rhoban's servo physics model. Treating a servo as an ideal torque source is the most common root cause of sim2real failure.
- Domain randomisation
Randomising physics during training so the policy holds across the real distribution. Not the same as adding observation noise.
- The 61-dimensional observation contract
Every policy shares one observation format, which is what lets the runtime hot-swap between walking, recovery and tricks.
- Policy hot-swapping
Switching between walking, recovery and trick policies at runtime, with no restart and no stop.
- Projected gravity: how the robot knows it fell
The only IMU quantity on the state stream. Upright is about [0, 0, -1], and the fall verdict is drawn from it.
- What sim2real is, and why it is the hard step
Moving a policy from simulation onto hardware. A policy that walks well in simulation and falls over on the robot is the normal outcome.
- Train and deploy
From MuJoCo training to a policy running on the robot, and the four things that decide whether it survives the move.
- Set up training and get your first policy
What microduck_rl needs, the four commands that matter, and what to do when you have no GPU.
- The sim2real recipe
The four things Microduck relies on to get from simulation onto hardware: BAM actuator physics, domain randomisation, backlash modelling, one shared observation contract.
- The training task list
Every trainable task in microduck_rl — id, terrain and purpose — including the roller family and the Backlash twins.
- BallKick — kicking a ball
Kicking a 70 mm, 15 g ball forward. The policy cannot see the ball.
- GroundPick — picking things up
Crouch and touch the ground with the beak tip, then return to standing. The A button.
- RollerCrouch — crouching while gliding
Dropping the centre of mass while keeping the glide.
- RollerSlope — gliding down slopes
Descending a slope on the wheels while staying in control.
- RollerStandUp — getting back onto the wheels
Standing up off the floor and back onto the wheels after a fall.
- Rollers — skating on velocity commands
The base policy for the passive wheels. The roller mode equivalent of the main walking task.
- Roulade — the forward roll
Rolling forward over the head and landing back on the feet. X / square, hold to chain.
- SitStand — sitting down and standing up
Commanded sit ↔ stand in one policy, gently, with the head still controllable.
- Spin — spinning in place
Spinning fast in place on the wheels.
- StandUp — getting up off the floor
Standing up from face-down, face-up or sitting, then holding the stand with body-pose control.
- Swizzle — the symmetric skating stroke
The classic swizzle: both feet push out and draw back in together.
- Velocity — walking on velocity commands
The main task: walking to velocity commands while taking head-pose commands. Train this one first.
- VelStand — walking and fall recovery together
Walking and fall recovery trained into one policy, which removes the handover at runtime.
- Background: what Hugging Face acquiring Pollen Robotics means
Microduck comes from a company Hugging Face owns. That relationship explains why its behaviours ship as shareable ONNX files.
- BAM — the servo actuator model
Rhoban's actuator identification framework. Microduck uses its M6 model for the Dynamixel XL330.
- microduck_rl — the training environments
Where every Microduck policy is trained: MuJoCo Warp with PPO, plus domain randomisation, the BAM actuator model and backlash simulation.
- Microduck Simulator — the browser simulator
The Microduck Sandbox on Hugging Face Spaces. No hardware needed.
- mjlab — the framework microduck_rl builds on
The GPU-parallel reinforcement learning framework built on MuJoCo Warp that Microduck's training sits on top of.