# RoboTwin2 Checkpoints

ACT, and pi0.5 single-task finetuning using B200 GPU on [RoboTwin2.0](https://github.com/TianxingChen/RoboTwin) dataset. 

## The policies were trained on the following Tasks:
- `place_phone_stand`
- `place_a2b_left`
- `move_can_pot`
- `handover_block`
- `put_bottles_dustbin`

## Data
- **Demonstrations:** 50 `demo_clean` episodes per task
- **Embodiment:** aloha-agilex (dual-arm)
- **Action dim:** 14 (6 DOF × 2 arms + 2 grippers)
- **Cameras:** `cam_high`, `cam_right_wrist`, `cam_left_wrist`

---

## ACT (Action Chunking Transformers)

### Architecture
| Param | Value |
|---|---|
| Backbone | ResNet-18 |
| Hidden dim | 512 |
| Feedforward dim | 3200 |
| Attention heads | 8 |
| Encoder layers | 4 |
| Decoder layers | 7 |
| Chunk size | 50 |
| KL weight | 10 |
| Action dim | 14 |
| Dropout | 0.1 |
| Parameters | ~83.9M |

### Training
| Param | Value |
|---|---|
| Batch size | 8 |
| Epochs | 6000 |
| Learning rate | 1e-5 |
| LR backbone | 1e-5 |
| Weight decay | 1e-4 |
| Optimizer | AdamW |
| Save freq | every 2000 epochs |

### Checkpoints
| Path | Seed | Val Loss |
|---|---|---|
| `ACT/act-place_phone_stand/demo_clean-50/` | 0 | — |
| `ACT/act-place_phone_stand-run2/demo_clean-50/` | 1 | 0.038 |
| `ACT/act-place_a2b_left/demo_clean-50/` | 0 | — |
| `ACT/act-place_a2b_left-run2/demo_clean-50/` | 1 | 0.059 |
| `ACT/act-move_can_pot/demo_clean-50/` | 0 | — |
| `ACT/act-move_can_pot-run2/demo_clean-50/` | 1 | 0.036 |
| `ACT/act-handover_block-run2/demo_clean-50/` | 1 | 0.030 |
| `ACT/act-put_bottles_dustbin-run2/demo_clean-50/` | 1 | 0.032 |

Each checkpoint directory contains:
- `policy_best.ckpt` — best validation loss checkpoint
- `policy_last.ckpt` — final epoch checkpoint
- `policy_epoch_{2000,4000,5000,6000}_seed_{0,1}.ckpt` — intermediate checkpoints
- `dataset_stats.pkl` — normalization statistics

---

## Pi0.5 LoRA (place_phone_stand only)

Fine-tuned from `gs://openpi-assets/checkpoints/pi05_base/params` using the [openpi](https://github.com/Physical-Intelligence/openpi) framework.

### Architecture
| Param | Value |
|---|---|
| Base model | Pi0.5 (3B params) |
| PaliGemma variant | `gemma_2b_lora` |
| Action expert variant | `gemma_300m_lora` |
| Fine-tuning method | LoRA |

### Training
| Param | Value |
|---|---|
| Batch size | 32 |
| Total steps | 20,000 (trained to 9,000) |
| Save interval | 200 steps |
| XLA memory fraction | 0.45 (64 GB pool on H200) |
| GPU | NVIDIA H200 (143 GB VRAM) |

### Checkpoints
| Path | Step |
|---|---|
| `pi05_lora/place_phone_stand/step_5000/` | 5,000 |
| `pi05_lora/place_phone_stand/step_9000/` | 9,000 |

---

## Environment
- **Framework:** [RoboTwin2.0](https://github.com/TianxingChen/RoboTwin)
- **Simulator:** SAPIEN with Vulkan rendering
- **GPU:** NVIDIA H200 SXM (143 GB VRAM)
- **CUDA:** 12.8