| configs | ||
| docs | ||
| .gitignore | ||
| README.md | ||
A1XY VLM / VLA study
Master's research plan: compare a modular VLM-based system with an end-to-end VLA policy on controlled tabletop object sorting using a real Galaxea A1X. ACT is a non-language baseline. No training pipeline, autonomous sorting implementation or experimental results exist here yet.
Current hardware: one A1X follower and one Star Arm 102-LD leader. Earlier two-arm social-video tests used another installation. Start with single-arm sorting; bimanual handover is conditional on later access to a second arm.
The proposal is summarized in the study protocol, based on the supplied
PREZKA.pdf and the owner's corrected home hardware inventory. The PDF remains a local reference
under references/local/ and is excluded from Git. Its authorship is not reassigned by this summary.
The proposal contains hypotheses, not measured conclusions.
Repository ownership
- a1xy-sdk: reusable wire codec and native USB transport.
- a1xy-teleop: camera control, web panel, motion lifecycle, calibration and simulation.
- lerobot-a1xy: planned A1X/Star Arm integration and recording.
- This repository: task definitions, policies, experiment configuration, evaluation and thesis work.
The modular pipeline proposes RGB-D perception, VLM task decisions, 3D grasp selection and classical motion planning. The learned pipeline proposes demonstrations recorded with the leader, followed by fine-tuning and evaluation. The leader and camera arrangement should keep human hands outside training frames; recordings must verify this.
configs/sorting-single-a1x.yaml is a protocol draft, not executable robot configuration. It includes
no hardware enable flags or borrowed calibration. Camera models, policy checkpoints, planner setup
and trial counts remain to be selected. There is no launch or training command to run yet.
The vendor email draft requests SDK access and protocol clarification. It has not been sent. Dataset, video and model storage stays outside Git; record hashes and manifests when data collection begins. Keep credentials and calibration local.