For a small robot, the gap between simulation and reality can be dominated not by the learning algorithm but by the servos: voltage, friction, load, and gear play all shape how a step actually lands. This project makes those hardware effects part of the training problem, turning sim-to-real transfer into an explicit engineering target rather than an afterthought.
What Sets It Apart
- Actuator fidelity is central. The training setup models the Dynamixel XL330 with voltage control, back-EMF, and load-dependent friction. That matters especially at this scale, where simplified ideal-PD motors can produce gaits that work in simulation but fail on the physical robot.
- Randomization targets physical uncertainty. Battery voltage, voltage sag, command delay, and friction vary across environments, while backlash variants add gear play to servo joints. Policies must therefore tolerate hardware variation, not just noisy observations.
- One observation contract supports policy switching. Walking, recovery, posture, and trick policies share a 61-dimensional observation layout. The robot runtime can swap among them without changing the interface, making a growing repertoire easier to deploy.
- The task suite goes beyond walking. It includes standing up, sit-to-stand transitions, ground-touching, ball kicking, forward rolls, and roller-skating behaviors, providing a practical testbed for learned whole-body control.
Who It’s For
Great fit if you work on reinforcement learning for physical robots and want to study actuator-aware simulation, domain randomization, or policy deployment on a compact biped. Look elsewhere if you need a CPU-only training workflow: the main training path uses MuJoCo Warp and requires a CUDA GPU, although jobs can be submitted to Hugging Face Jobs.
