a2a_real
A2A flow-matching checkpoints for real-robot tasks. This repository was renamed from
SeanLiu0272/a2a_fruitV3_40; the original fruit artifacts remain at the repository root.
Nut/washer and pen/cabinet checkpoints
Added 2026-11-11. Both runs completed 500 training epochs. Each run directory contains
all five saved checkpoints under checkpoints/.
| Run directory | Saved checkpoint files | Best epoch | Best validation action MSE |
|---|---|---|---|
nut_washer_v3_N77_20261108_192319 |
200.ckpt, 300.ckpt, 400.ckpt, 500.ckpt, best.ckpt |
10 | 0.005428999920508691 |
pen_cabinet_N67_20261108_072011 |
250.ckpt, 300.ckpt, 400.ckpt, 500.ckpt, best.ckpt |
10 | 0.003563778771619712 |
Checkpoint filenames and file contents are preserved from the source runs.
Original fruitV3_40 artifacts
A2A flow-matching policy (a2a_noise, A2ANoiseImagePolicy, 47.9M params) trained on the real-robot
fruitV3_40 dataset (40 teleop episodes, 30 Hz, head cam agentview_rgb + wrist cam
robot0_eye_in_hand_rgb, 128x128; 7-d action). Code: https://github.com/seanliu7081/A2A_Flow_Matching
(roboverse_learn/il/il_run_real_robot.sh, see roboverse_learn/il/README_real_robot.md).
Snapshot taken mid-training (2026-11-08): epochs 0-156 of a planned 500. The original fruit artifacts are a training snapshot; the epoch-250 checkpoint was added separately as noted below.
| file | contents |
|---|---|
checkpoints/best.ckpt |
lowest val_action_mse so far (epoch 25, 0.00887); see best.json |
checkpoints/50.ckpt, 100.ckpt, 150.ckpt |
periodic checkpoints (model + EMA + normalizer) |
checkpoints/250.ckpt |
periodic checkpoint at epoch 250, added 2026-11-09 (logs/latest.ckpt remain from the epoch-156 snapshot) |
checkpoints/latest.ckpt |
full training state after epoch 156 (incl. optimizer/RNG), resumable with --resume |
split.json, run_info.json, .hydra/, launch_*.txt |
seed-42 10% episode split, config, launch command |
console.log, train.log, logs.json.txt, latent_viz/, wandb/ |
training logs (W&B run xcmtu4i1; binary .wandb log omitted, e-mail redacted) |
Load with torch.load(path, map_location="cpu", pickle_module=dill, weights_only=False), then call
.to(device) on the normalizer after load_state_dict. MANIFEST.sha256 lists SHA-256 checksums for repository files, excluding the manifest itself.