Two small robot arms from a Waveshare SO-ARM101 kit, built from 3D-printed parts: a leader arm that a person moves by hand, and a follower arm that copies it. This is how the demonstrations will be recorded. So far, every motion is controlled by a person; no controller has been trained yet.

The planned task: the follower arm pushes a puck through a gate, while a controlled sideways push and a controlled delay cause mistakes it can still recover from.

Printed arm parts, servos, power supplies, a USB hub, cameras, clamps and a screwdriver set laid out on a table
Six small paper labels marked F1 to F6 for the follower arm servos
Two servo bus driver boards, one labelled LEADER and one labelled FOLLOWER

Context

Primary references

  1. Zhao et al. (2023): Learning fine-grained bimanual manipulation with low-cost hardware (ACT)
  2. Ross, Gordon & Bagnell (2011): DAgger: reduction of imitation learning to no-regret online learning
  3. Ramstedt & Pal (2019): Real-time reinforcement learning: acting under one-step delay
  4. Cadene et al.: LeRobot (open-source robot learning tooling)

Novelty scope: Controlling a robot by hand, learning from demonstrations, and compensating for delays are established techniques. The planned contribution is a careful comparison of recovery data and delay-aware execution on affordable hardware.