Kyoto University NF: Imitation Learning Demo with a Self-Built ALOHA

Overview An ALOHA costs millions of yen, which is not a student budget. So we built ILOHA — an ALOHA compatible bimanual manipulator — designed the data collection pipeline around a Meta Quest 3, trained an ACT policy on what we recorded, and demonstrated it live at the Kyoto University November Festival (NF). The robot: a hybrid actuator layout The cost of an ALOHA is dominated by its actuators, but not every joint needs the same one: Wrist / distal joints — Dynamixel. These sit inside the camera's field of view, and how they look genuinely matters to a vision based policy, so we kept the standard part here. Base / proximal joints — QDD motors (Robstride). These are barely visible in the observations, so a much cheaper actuator does the job. The result stays faithful to the reference design where the policy can see it, and saves money where it cannot. Data collection: teleoperation from a Quest 3 Rather than build a leader arm, we teleoperate from a Meta Quest 3: Controller pose → inverse kinematics → target joint angles, sent to the robot through a PC. A RealSense camera records images synchronized with the robot state, in the format the LeRobot library expects — so recorded data flows into training with no conversion step. Getting this stable took more work than the IK itself: Gimbal lock near singular poses was handled with a singularity avoidance step on the Quest side. Runaway motion was prevented on the PC side with angle unwrapping, per step relative motion limits, and exponential smoothing of the commands. Packet loss turned out to be the real enemy. Swapping the…