EdgeArm Run102 model card
License: Apache-2.0. Model repository: YuxuanGong/EdgeArm-Run102.
Four exact checkpoint files are released with SHA-256 checksums:
| File | Role |
|---|---|
checkpoint.pt |
Run82 update-500 learned spatial memory, including its observation backbone |
control.pt |
Run100 round-1 action MLP |
keypoint.pt |
Run101 update-1200 current-frame keypoints |
vision.pt |
Historical compatible vision object loaded by the frozen interface; not an extra active ensemble |
These are the exact frozen research checkpoints; inspect them using PyTorch's
weights_only=True, never load arbitrary untrusted pickle files. Source hashes and
model hashes are separate: portable release wrappers are new code, not part of the
original recorded 72 episodes.
Observations and actions
Wrist RGB + reported joints/velocities/FK + causal action/joint history + camera calibration + explicit seven-color instruction parsing. The controller receives 114 proprioceptive/history values and four estimated block/target XY values. Its six normalized joint command outputs pass through the existing command contract and joint/workspace/camera projection. No ground-truth object coordinates, teacher actions, route IDs, or future frames are deployment inputs.
Evaluation
Nominal MuJoCo simulation, task-conditioned CONTACT_TRANSPORT_HOLD initial pose,
nine equally weighted routes, eight held-out scene groups, 72 complete episodes.
51 succeeded, 20 timed out, one went out of bounds, zero hard-contact failures.
All successes met the original coverage, speed, and consecutive three-second hold
criterion. There is a fixed 220-step observation program inside the 900-step total.
Wilson approximate 95% interval: 59.49–80.06%; scene-group bootstrap: 63.89–77.78%. No real-robot, exact-Home, broad language, or fully randomized-environment result is claimed. The weights are a research release, not hardware safety certification. Training data remains private by owner choice. Production admission remains false.