Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# RoboTwin2 Checkpoints
|
| 2 |
+
|
| 3 |
+
Trained on [RoboTwin2.0](https://github.com/TianxingChen/RoboTwin) using the [TianxingChen/RoboTwin2.0](https://huggingface.co/datasets/TianxingChen/RoboTwin2.0) dataset.
|
| 4 |
+
|
| 5 |
+
## Tasks
|
| 6 |
+
- `place_phone_stand`
|
| 7 |
+
- `place_a2b_left`
|
| 8 |
+
- `move_can_pot`
|
| 9 |
+
- `handover_block`
|
| 10 |
+
- `put_bottles_dustbin`
|
| 11 |
+
|
| 12 |
+
## Data
|
| 13 |
+
- **Demonstrations:** 50 `demo_clean` episodes per task
|
| 14 |
+
- **Embodiment:** aloha-agilex (dual-arm)
|
| 15 |
+
- **Action dim:** 14 (6 DOF × 2 arms + 2 grippers)
|
| 16 |
+
- **Cameras:** `cam_high`, `cam_right_wrist`, `cam_left_wrist`
|
| 17 |
+
|
| 18 |
+
---
|
| 19 |
+
|
| 20 |
+
## ACT (Action Chunking Transformers)
|
| 21 |
+
|
| 22 |
+
### Architecture
|
| 23 |
+
| Param | Value |
|
| 24 |
+
|---|---|
|
| 25 |
+
| Backbone | ResNet-18 |
|
| 26 |
+
| Hidden dim | 512 |
|
| 27 |
+
| Feedforward dim | 3200 |
|
| 28 |
+
| Attention heads | 8 |
|
| 29 |
+
| Encoder layers | 4 |
|
| 30 |
+
| Decoder layers | 7 |
|
| 31 |
+
| Chunk size | 50 |
|
| 32 |
+
| KL weight | 10 |
|
| 33 |
+
| Action dim | 14 |
|
| 34 |
+
| Dropout | 0.1 |
|
| 35 |
+
| Parameters | ~83.9M |
|
| 36 |
+
|
| 37 |
+
### Training
|
| 38 |
+
| Param | Value |
|
| 39 |
+
|---|---|
|
| 40 |
+
| Batch size | 8 |
|
| 41 |
+
| Epochs | 6000 |
|
| 42 |
+
| Learning rate | 1e-5 |
|
| 43 |
+
| LR backbone | 1e-5 |
|
| 44 |
+
| Weight decay | 1e-4 |
|
| 45 |
+
| Optimizer | AdamW |
|
| 46 |
+
| Save freq | every 2000 epochs |
|
| 47 |
+
|
| 48 |
+
### Checkpoints
|
| 49 |
+
| Path | Seed | Val Loss |
|
| 50 |
+
|---|---|---|
|
| 51 |
+
| `ACT/act-place_phone_stand/demo_clean-50/` | 0 | — |
|
| 52 |
+
| `ACT/act-place_phone_stand-run2/demo_clean-50/` | 1 | 0.038 |
|
| 53 |
+
| `ACT/act-place_a2b_left/demo_clean-50/` | 0 | — |
|
| 54 |
+
| `ACT/act-place_a2b_left-run2/demo_clean-50/` | 1 | 0.059 |
|
| 55 |
+
| `ACT/act-move_can_pot/demo_clean-50/` | 0 | — |
|
| 56 |
+
| `ACT/act-move_can_pot-run2/demo_clean-50/` | 1 | 0.036 |
|
| 57 |
+
| `ACT/act-handover_block-run2/demo_clean-50/` | 1 | 0.030 |
|
| 58 |
+
| `ACT/act-put_bottles_dustbin-run2/demo_clean-50/` | 1 | 0.032 |
|
| 59 |
+
|
| 60 |
+
Each checkpoint directory contains:
|
| 61 |
+
- `policy_best.ckpt` — best validation loss checkpoint
|
| 62 |
+
- `policy_last.ckpt` — final epoch checkpoint
|
| 63 |
+
- `policy_epoch_{2000,4000,5000,6000}_seed_{0,1}.ckpt` — intermediate checkpoints
|
| 64 |
+
- `dataset_stats.pkl` — normalization statistics
|
| 65 |
+
|
| 66 |
+
---
|
| 67 |
+
|
| 68 |
+
## Pi0.5 LoRA (place_phone_stand only)
|
| 69 |
+
|
| 70 |
+
Fine-tuned from `gs://openpi-assets/checkpoints/pi05_base/params` using the [openpi](https://github.com/Physical-Intelligence/openpi) framework.
|
| 71 |
+
|
| 72 |
+
### Architecture
|
| 73 |
+
| Param | Value |
|
| 74 |
+
|---|---|
|
| 75 |
+
| Base model | Pi0.5 (3B params) |
|
| 76 |
+
| PaliGemma variant | `gemma_2b_lora` |
|
| 77 |
+
| Action expert variant | `gemma_300m_lora` |
|
| 78 |
+
| Fine-tuning method | LoRA |
|
| 79 |
+
|
| 80 |
+
### Training
|
| 81 |
+
| Param | Value |
|
| 82 |
+
|---|---|
|
| 83 |
+
| Batch size | 32 |
|
| 84 |
+
| Total steps | 20,000 (trained to 9,000) |
|
| 85 |
+
| Save interval | 200 steps |
|
| 86 |
+
| XLA memory fraction | 0.45 (64 GB pool on H200) |
|
| 87 |
+
| GPU | NVIDIA H200 (143 GB VRAM) |
|
| 88 |
+
|
| 89 |
+
### Checkpoints
|
| 90 |
+
| Path | Step |
|
| 91 |
+
|---|---|
|
| 92 |
+
| `pi05_lora/place_phone_stand/step_5000/` | 5,000 |
|
| 93 |
+
| `pi05_lora/place_phone_stand/step_9000/` | 9,000 |
|
| 94 |
+
|
| 95 |
+
---
|
| 96 |
+
|
| 97 |
+
## Environment
|
| 98 |
+
- **Framework:** [RoboTwin2.0](https://github.com/TianxingChen/RoboTwin)
|
| 99 |
+
- **Simulator:** SAPIEN with Vulkan rendering
|
| 100 |
+
- **GPU:** NVIDIA H200 SXM (143 GB VRAM)
|
| 101 |
+
- **CUDA:** 12.8
|