File size: 2,518 Bytes
1e71a55
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
## 1. Example Code Guide

This document provides a minimal end-to-end example for training, evaluating, and submitting a policy.

## 2. RL for Locomotion

### 2.1 Train a PPO Policy (Example)

The baseline workflow references: https://github.com/fan-ziqi/robot_lab

Run the following command from the repository root:

```bash
python scripts/rsl_rl/train.py --task ATEC-Isaac-Velocity-Flat-Unitree-B2-v0 --headless --video
```

On an **NVIDIA RTX 5090**, this example typically takes around **90 minutes**.

Actual training time depends on driver/runtime version, CPU performance, and current GPU load.

### 2.2 Evaluate the Trained Policy

After training, evaluate with:

```bash
python scripts/rsl_rl/play.py --task ATEC-Isaac-Velocity-Flat-Unitree-B2-v0
```

This loads the trained checkpoint and runs rollout in the same task setting.

### 2.3 Test Locally

The file `demo/solution.py` is the only entrance for locally testing and online submission.
Use the test command:

```bash
cd ATEC2026_Simulation_Challenge
python scripts/play_atec_task.py --task ATEC-TaskA-B2Piper --enable_cameras
```

Notes:

- `--task` selects the arena and robot. See the Environment Matrix in `readme.md`.
- Use `--debug` to print runtime status and score.

Pretrained baseline checkpoint:

- `./atec_robot_model/baseline/unitree_b2_flat/policy.pt`
This checkpoint path can be modified in `demo/solution.py`.

![baseline demo](./doc/baseline.gif)

## 3. IL for Manipoulation

This section follows the core idea of ACT (Action Chunking with Transformers).

Reference implementation: https://github.com/tonyzhaozh/act

### 3.1 Collect Demonstrations

Collect expert trajectories for Task E:

```bash
python scripts/act/collect_demos_task_e.py --pick_objects 3 --num_demos 100 --headless --enable_cameras --save_images
```

Filter out near-zero actions from the collected dataset:

```bash
python scripts/act/filter_demos.py \
	--input datasets/atec_task_e/trajectory.hdf5 \
	--output datasets/atec_task_e/trajectory_filtered.hdf5 \
	--threshold 0.001
```

### 3.2 Train ACT Policy

Run ACT baseline training from the `scripts/act` directory:

```bash
cd scripts/act
bash baseline.sh
```

### 3.3 Run the Trained Policy

Use the test command:

```bash
python scripts/play_atec_task.py --task ATEC-TaskE-Piper --enable_cameras
```

Note: `./atec_robot_model/baseline/act/policy.pt` is the provided baseline checkpoint. You can replace it with your own trained policy path in `demo/solution`.

![baseline act demo](./doc/baseline_act.gif)