Training complete: 200/200 epochs, 37.97h, zero errors, all 400 progress bars ran their full step count (no DataLoader starvation).
Adds checkpoint_final.pth (epoch 199) alongside checkpoint_epoch100.pth. Neither is Isaac Gym validated yet, and the training curves do not rank them: epoch 199 wins on train acc (0.7283) and val acc (0.6923), epoch 100 wins on val loss (1.7518). Evaluate both.
README.md is a full evaluation guide: env setup, keypoint staging, the generate_grasps_for_obj.py hardcoded-filename trap, a working bridge script for the gen_grasps.pt -> run_grasp_test.py format gap, and Isaac Gym's separate py3.8+CUDA requirement. INFO.md updated to reflect completed training.
Add geomatch_v3 PREVIEW: epoch-100 checkpoint trained on real keypoints
Mid-training checkpoint (epoch 100/200), NOT Isaac Gym validated.
First run trained on the real robot_keypoints.json rather than the centroid-generated stand-in used by geomatch_v2_repro. Includes both keypoint JSONs, which were never released upstream (deepmind/geomatch issue #1).
See geomatch_v3/INFO.md for verification, contact-label comparison, and caveats.
Upload geomatch_v2_repro/EVAL_INFO.md with huggingface_hub
