demo-collection

Collect robot demonstration data for reinforcement learning by using kinesthetic teaching, teach pendant, or scripted demos, then processing and exporting the results in RLDS format for model fine-tuning.

The RLDS format structures reinforcement learning demonstration data. You need it when exporting collected robot demonstrations to ensure compatibility with downstream reinforcement learning frameworks.

Set up a Doosan e0509 robot and a Realsense D455f camera for depth sensing to enable kinesthetic teaching. These hardware components capture and process the physical demonstrations for fine-tuning.

Yes, the demonstration collection method supports scripted demos and teach pendant input, alongside kinesthetic teaching, providing multiple scenarios to automate and validate high-quality data collection.

Yes, you can use collected demonstration data for VLA fine-tuning. The collected and validated data exports in RLDS format, making it ideal for Phase 6a demonstration collection in VLA environments.

Kinesthetic teaching requires physical manipulation of the Doosan e0509 robot and accurate Realsense D455f depth sensing, meaning hardware setup and physical presence are mandatory limitations for this approach.