HIL-SERL (human-in-the-loop RL)

How it works

Start with a handful of demonstrations and a simple learned classifier that says whether the task succeeded. The robot then practices on its own with reinforcement learning. Whenever it is about to fail, a person steps in with a 3D mouse and shows the right move; those corrections go into the training data. Within a couple of hours the policy stops needing corrections and reaches 100% on precise tasks like inserting a RAM stick, because it learned from its own mistakes, not just from copying.

What it needs

A real robot you can let practice safely (resets matter), wrist and side cameras, a SpaceMouse for interventions, a person present for 1–6 hours per task, and a GPU. Works with torque- or velocity-controlled arms.

Strengths

• 100% success on all 13 real tasks in the paper, including precision insertion and assembly
• Trains in hours not days
• Policies are reactive and recover from perturbations

Limits

• One policy per task; no generalization claim
• Needs a human present during training
• Reset and safety engineering is on you