Imitation-Learning-Agent/README.txt at main · N-Raghav/Imitation-Learning-Agent · GitHub

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
Project 4 — Imitation Learning Submission
==========================================

Environment
-----------
All training was run inside an Apptainer (Singularity) container (robomimic.sif) on
the WPI Turing cluster (SLURM, 1x GPU per job, CUDA 12.9).

Framework: robomimic (custom fork in repo)
Task:      Stack with UR5e arm (converted.hdf5, 143 train demos with RGB images)

Each method was run twice:
  Run 1 — default / first-attempt configuration
  Run 2 — improved configuration (ablation study)

Training Configs  (training_configs/)
--------------------------------------
bc_run1.json          BC-Transformer v2 Run 1 (context 3, batch 100, no step supervision)
bc_run2.json          BC-Transformer v2 Run 2 (context 10, batch 64, supervise_all_steps)
diff_run1.json        Diffusion Policy Run 1 (DDPM, 100 inference steps, 2000 epochs)
diff_run2.json        Diffusion Policy Run 2 (DDIM, 10 inference steps, 1000 epochs)

Checkpoints  (checkpoints/)
-----------------------------
bc_run1_model.pth     BC-Transformer v2 Run 1, epoch 950, Success Rate = 30%
bc_run2_model.pth     BC-Transformer v2 Run 2, epoch 950, Success Rate = 100%
diff_run1_model.pth   Diffusion Policy Run 1 (DDPM), epoch 650, Success Rate = 80%
diff_run2_model.pth   Diffusion Policy Run 2 (DDIM), epoch 350, Success Rate = 80%

Logs  (logs/)
--------------
bc_run1_training_log.txt       Full training log for BC Run 1
bc_run2_training_log.txt       Full training log for BC Run 2
diff_run1_training_log.txt     Full training log for Diffusion Policy Run 1
diff_run2_training_log.txt     Full training log for Diffusion Policy Run 2

Tensorboard  (tensorboard/)
-----------------------------
bc_run1/              TFEvents for BC-Transformer Run 1
bc_run2/              TFEvents for BC-Transformer Run 2
diff_run1/            TFEvents for Diffusion Policy Run 1
diff_run2/            TFEvents for Diffusion Policy Run 2

To view: tensorboard --logdir tensorboard/

Videos  (videos/)
------------------
bc_run1_rollout.mp4        BC-Transformer Run 1 at epoch 950 (30% success)
bc_run2_rollout.mp4        BC-Transformer Run 2 at epoch 950 (100% success)
diff_run1_rollout.mp4      Diffusion Policy Run 1 (DDPM) at epoch 650 (80% success)
diff_run2_rollout.mp4      Diffusion Policy Run 2 (DDIM) at epoch 350 (80% success)

Scripts  (scripts/)
--------------------
bc_transformer_v2_job.sh    SLURM job script for BC Run 2 training
diffusion_imp_job.sh        SLURM job script for Diffusion Policy Run 2 training
demo_sweep.sh               Self-submitting SLURM array job for demo-count sweep
                            (trains BC on 5/10/20/50 demos, 4 tasks total)
plot_demo_sweep.py          Parses slurm_op/ logs and generates success-rate plots
merge_demos.py              Merges multiple demo HDF5 files into one dataset
merge_two_datasets.py       Merges exactly two HDF5 datasets
patch_env_args.py           Patches robomimic env args for UR5e compatibility