Imitation Learning for Micro-Scale Manipulation: A Diffusion Policy Approach on an Open-Source Delta Micro-Manipulator

Micro-scale manipulation below the millimetre range remains limited by the cost and closed design of commercial platforms, which makes results hard to reproduce and slows methodological progress. At this scale manipulation is still performed almost entirely by teleoperation or by hand-designed controllers, and the learning-based methods that now let robots acquire manipulation skills from demonstrations have so far been developed and tested only at the macro scale. This thesis builds an open-source micro-manipulator and trains visuomotor policies that perform contact manipulation tasks autonomously at the sub-millimetre scale, among the first demonstrations of learned autonomous manipulation at this scale. The platform is a 3-DOF delta-style parallel micro-manipulator (Open Micro-Manipulator v4.0 by 0x23/Diffraction Limited), actuated through PWM-driven H-bridge drivers and 21-bit magnetic encoders whose magnetic gearing gives an end-effector resolution of about 50 nm. Its deviations from the reference design are documented together with their consequences for accuracy. Two tasks are addressed. The first is a micro-scale version of Push-T, a standard benchmark in robotic manipulation, with a working area under 8 millimetre across, for which 329 demonstrations were recorded with synchronised microscope and encoder logging. The second is a colour-separation task defined in this work, in which a blue PLA blob about 1 millimetre across must be pushed out of a group of green ones and into a target zone, so that the policy has to choose where to make contact among several objects; 111 demonstrations were collected for it. A Diffusion Policy was trained on the recorded demonstrations and compared with an Action Chunking with Transformers (ACT) policy, on the Push-T task across architecture, dataset size and training budget, and on the separation task across architecture alone. On the Push-T task, the best results came from a Diffusion Policy trained on all 329 demonstrations. It solves 19 of 24 freely placed trials and 11 of 12 fixed initial conditions, reaches a median placement error of 0.068 mm against the operator's 0.055 mm, and successfully brings the object back to the goal after external disturbances of more than a millimetre. The amount of data matters more than the architecture: performance rises sharply between 200 and 329 demonstrations, ACT is ahead on the smaller subsets, and on the full dataset its failures end much farther from the goal. The remaining failures stem from the demonstrations rather than from the hardware: 62% of them leave the object in place but on the wrong arm of its three-fold symmetric shape, a correction the operator almost never had to demonstrate. On the separation task the Diffusion Policy is clearly the better model, placing the blob in the target zone more often and more precisely than ACT and never performing worse from any of the shared starting positions. The thesis shows that imitation learning can carry autonomous contact manipulation below the millimetre on low-cost hardware, with placement precision that approaches the operator's own, and contributes an open, reproducible platform, dataset and methodology for continuing this line of work.