AW8N stage-A tokenizers for C3 (RoboTwin / RoboDojo)
These are the two stage-A AW8N dynamic-code tokenizers used to label the 12 codes of the C3 policy.
The code that uses them, the training recipe and the step-by-step instructions are in the GitHub repository
yangzaijia/dynarobot-c3-train, which is private and needs access.
| file | bytes | sha256 | what it is |
|---|---|---|---|
robotwin/aw8n_trRGB_ckpt_0100000.pt |
2,306,965,982 | 51268003931f46c9a11211d357b7d3beb6e97b5500a8a6b866a7de84d18901be |
RoboTwin AW8N, step 100000 (S1 50k + S3 50k, ฮ16โ64, raw weights, RGB input); labels use d50 (delta index 34) |
robodojo/aw8n_d23_27_ckpt_0150000.pt |
1,729,705,861 | a2f689ae6d22c5493c7b162c5ac5fa5f416d31c37c074e673ea568ad7b82c3b1 |
RoboDojo AW8N, global step 150000 (S1 100k + S3 50k, ฮ23โ27); labels use d25 (delta index 2) |
Each file is a torch checkpoint dict with the keys model, step, quantized, args, โฆ. Only model is needed: it is loaded
into AW8N(**model_args(n_delta)) from aw8n/unit_model.py. The frozen C3 codebook is unit_vq.directions, shape (3, 64, 32).
Both files are byte-identical copies of the originals in April-Yang/dynarobotic:
dyntok/AW8N_robotwin_official_d56_64_mix20_d16_55_S1_50k_S3_50k_eff256_ddp2_m128a1_trRGB_viz_runhao_20261029_yzj/ckpt_0100000.ptdyntok/AW8N_robodojo_d23_27_raw100k_S3_50k_eff256_c4g2_20261017_yzj/ckpt_0150000.pt
They were copied and their sha256 verified on both ends on 2026-11-13.