Download astra/CoverPattern-v1/handoff.md from Linslab/ManiLadder_Models: direct link, hf CLI and curl.
- Browser
- Download file 2.19 kB
-
https://huggingface.co/Linslab/ManiLadder_Models/resolve/main/astra/CoverPattern-v1/handoff.md
- Command line
-
hf download hf://Linslab/ManiLadder_Models/astra/CoverPattern-v1/handoff.md
-
curl -L -o handoff.md https://huggingface.co/Linslab/ManiLadder_Models/resolve/main/astra/CoverPattern-v1/handoff.md
Agent-D20-E20-Compiled / CoverPattern-v1
Entrypoints are controller.py reset_episode() and act(obs). Keep paths.npz beside it. Dependencies: NumPy, OpenCV and Python standard library. No neural weights, network, model service, simulator imports or runtime development tools are used. The runner adds verified demo provenance and the common environment lock.
Input is legal RGB top_camera and proprioception agent/qpos; optional step_id enforces the 1500-step horizon. Output is one float32 length-22 action in native normalized order, or None. The runner owns the initial charged hold and episode lifecycle. The program cannot reset the environment or assert task success.
Initial colored rectangles select one of 20 authorized reference paths by position and yaw. Pixel displacement calibration adjusts pickups within 45 mm per axis, followed by a 30 mm north pickup offset, and destinations within 60 mm. Analytical Panda forward kinematics supplies relative Cartesian feedback; absolute table-to-root translation is not assumed. Episode-5 isolated legal probes establish the hand action/qpos permutation.
The controller captures finger posture at each drag onset and holds it through transport. Each cloth gets at most 64 image-guided pre-release correction steps. Release opens at gain .35 for 80 steps, combines a heuristic opening/withdrawal feedforward with visible cloth-center feedback, and then clears vertically for at most 40 steps. Release XY travel is bounded. Achieved displacements taper into the return path. Free-space path progression accelerates while grasp preparation and transport retain reference cadence. Pose waits total at most 120; the native horizon takes precedence over all internal timing. There is no full regrasp recovery.
All official results come from native API returns. Only if absent may legal images support the preregistered visible rule, with ambiguity false. Independent rehearsal receipts, exact hashes, full historical analysis, actions and images are retained outside this runtime bundle under exploration/ and demo_analysis.md. Candidate files are locked before final independent repetitions; later results belong in the external audit.