Views
No views yet
[object_start_xyz, object_goal_xyz][x, y, z, qx, qy, qz, qw, gripper_closed]| 项目 | 值 |
|---|---|
| objective | conditional Flow Matching |
| backbone | DiT |
| horizon | 256 |
| state dimension | 8 |
| condition dimension | 6 |
| hidden size | 256 |
| Transformer depth | 8 |
| attention heads | 8 |
| MLP ratio | 4 |
| training steps | 2300 |
| batch size | 16 |
| learning rate | 3e-4 |
| learning-rate warmup | 0 steps |
| EMA decay | 0.995 |
| seed | 7 |
| checkpoint | checkpoint_best.pt |
| best checkpoint step | 1800 |
| best validation loss | 0.103506 |
Yi-29/franka-plannerstart_* 和 goal_* 是物体坐标;observation.state 是实际测得的末端轨迹。duration 存在 environment state 的第 7 个值中。| sampler | candidates | ADE | FDE | orientation error | gripper accuracy |
|---|---|---|---|---|---|
| NFE=1 | 1 | 5.62 cm | 1.49 cm | 4.32° | 95.96% |
| NFE=1 | 8 | 4.47 cm | 0.85 cm | 4.18° | 97.01% |
| NFE=5 | 8 | 3.81 cm | 1.30 cm | 4.68° | 97.33% |
| NFE=10 | 8 | 3.87 cm | 1.29 cm | 4.82° | 97.66% |
1checkpoint_best.pt
2config.json
3normalization.json.pt 格式,应使用本项目的 imf_training 推理脚本,不要直接交给 lerobot-train 或当作 ACT policy 加载。