| --- |
| license: apache-2.0 |
| tags: |
| - robotics |
| - imitation-learning |
| - ACT |
| - dexterous-manipulation |
| - dexflow |
| library_name: pytorch |
| --- |
| |
| # DexFlow Stage-2 Checkpoints |
|
|
| Human→Robot DexFlow Stage-2 (Arena ACT + articulated HandFlow/ObjectFlow aux) checkpoints for Inspire-G1 tasks. |
|
|
| Org: [Humantwin](https://huggingface.co/Humantwin) |
|
|
| ## Wave-2 P0 (2026-09-09) — recommended for bottle_upright real-robot |
| |
| Same 10-ep **bottle_upright right crop cam30** robot set. P0-Ours additionally inits from Manus **plastic_bottle** Stage-1 (14 ep); S2 does not mix human RGB. |
| |
| | ID | File | Init | S2 loss | Open-loop body / hand / J2 | |
| |----|------|------|---------|----------------------------| |
| | **P0-B0** | `p0s2_b0_bottle_cam30_crop/stage2.ckpt` | ImageNet | ACT only | **0.0085 / 4.69 / 1.49** | |
| | **P0-Ours** | `p0s2_ours_p0b_hf_bottle_cam30_crop/stage2.ckpt` | plastic S1 | ACT + 0.25 HF | **0.0083 / 4.83 / 1.72** | |
| |
| Load: |
| |
| ```python |
| from dexflow.backbone.dexflow_act import DexFlowACTPolicy |
| model = DexFlowACTPolicy.from_checkpoint("p0s2_b0_bottle_cam30_crop/stage2.ckpt", stage="deploy") |
| model = model.strip_aux_for_deploy() |
| ``` |
| |
| Deploy: publish 30/30, infer interval ≥1.05 s, strip aux, MEAN_STD `norm_stats.json` in the same folder. |
| |
| --- |
| |
| ## bottle_upright (right, crop cam30) — older 15k aux run |
| |
| | Item | Value | |
| |---|---| |
| | File | `bottle_upright_right_crop/step_15000.ckpt` | |
| | Steps | 15000 / 30000 | |
| | Loss | L1 + 10·KL + 0.25·L_aux (HF+OF, contact off) | |
| | Init | Multi-task Stage-1 stage1_step06000 visual (+aux) | |
| | Data | bottle_upright right crop @ camera_30hz, 10 episodes | |
| | Norm | `bottle_upright_right_crop/norm_stats.json` (MEAN_STD) | |
| |
| ### Open-loop @ step 15000 (train-set, chunk step-0) |
| |
| | Metric | Value | |
| |---|---:| |
| | body_rmse | 0.0134 rad | |
| | hand_mae | 9.62 (0–1000) | |
| | action_mse | 74.5 | |
| |
| See `bottle_upright_right_crop/openloop_step15000/`. |
| |
| ### Load (DexFlow) |
| |
| ```python |
| import torch |
| ckpt = torch.load("bottle_upright_right_crop/step_15000.ckpt", map_location="cpu") |
| ``` |
| |
| ## Layout |
| |
| ```text |
| p0s2_b0_bottle_cam30_crop/ # Wave-2 P0-B0 @30k |
| stage2.ckpt |
| norm_stats.json |
| p0s2_ours_p0b_hf_bottle_cam30_crop/ # Wave-2 P0-Ours @30k |
| stage2.ckpt |
| norm_stats.json |
| bottle_upright_right_crop/ |
| step_15000.ckpt |
| norm_stats.json |
| openloop_step15000/ |
| ``` |
| |
| --- |
| |
| ## Core4 current-state MLP G2 @ 25k (2026-09-11) |
| |
| Four-task DexFlowMLP Stage-2 (action-only L1). Same 10 robot episodes / task. Same-set open-loop, not held-out SR. |
| |
| | Task | File | nRMSE41 | body RMSE | hand MAE | |
| |---|---|---:|---:|---:| |
| | plastic / bottle_upright | `core4_mlp_current_state_g2_step25000/bottle_upright_right_crop/stage2_step25000.ckpt` | 0.050 | 0.0050 | 2.74 | |
| | pill | `core4_mlp_current_state_g2_step25000/pill_bottle_right_crop/stage2_step25000.ckpt` | 0.067 | 0.0059 | 4.16 | |
| | open_box_lid | `core4_mlp_current_state_g2_step25000/open_box_lid_both_crop/stage2_step25000.ckpt` | 0.054 | 0.0089 | 3.01 | |
| | realsense | `core4_mlp_current_state_g2_step25000/realsense_box_into_box_right_crop/stage2_step25000.ckpt` | 0.063 | 0.0062 | 3.63 | |
|
|
| Load: `DexFlowMLPPolicy.from_checkpoint(..., stage="deploy")` + same-folder `norm_stats.json`. |
|
|
|
|
