File size: 8,902 Bytes
7bb6581 9213391 7bb6581 9213391 7bb6581 9213391 7bb6581 9213391 7bb6581 5bb8179 7bb6581 5bb8179 7bb6581 5bb8179 7bb6581 5bb8179 7bb6581 5bb8179 7bb6581 | 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 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 | # UR5 Sim2Real: Data Collection & Diffusion Policy Training
This repo contains the scripts used to collect simulation demonstrations and train Diffusion Policy (DP) checkpoints for 6 robot manipulation tasks on NVIDIA L40s cluster nodes, as part of the UR5 sim2real pipeline.
## Overview
| Task | Demos | DP Checkpoint | Notes |
|------|-------|---------------|-------|
| `battery_v1` | 50 | `1027_sim_ckpt/battery_v1/latest.ckpt` | ✅ |
| `ring_v1` | 50 | `1027_sim_ckpt/ring_v1/latest.ckpt` | ✅ |
| `package_d5_v1` | 50 | `1027_sim_ckpt/package_d5_v1/latest.ckpt` | ✅ |
| `ringpeg_dart_v1` | 100 | `1027_sim_ckpt/ringpeg_dart_v1/latest.ckpt` | ✅ |
| `blb_v1` | 47 | `1027_sim_ckpt/blb_v1/latest.ckpt` | ✅ (47 due to GPU crash) |
| `plug_v1` | 50 | `1027_sim_ckpt/plug_v1/latest.ckpt` | ✅ |
| `toaster_lever` | — | — | ❌ Skipped: USD crashes Isaac Sim 5 |
| `plug1ramp_v1` | 50 (to collect) | — | 🆕 2026-10-31, `yp_task_1031.sh` |
| `battery2_v1` | 50 (to collect) | — | 🆕 2026-10-31, `yp_task_1031.sh` |
| `eraser1_v1` | 50 (to collect) | — | 🆕 2026-10-31, `yp_task_1031.sh` |
Demo zarr + collect_meta are in `1027_sim_demo/<task>/` on HuggingFace `yqi19/dagger`.
---
## Repository Layout
```
1027_run/ # Original run (2x parallel collection)
tasks/
yp_task.sh # battery / ring / plug collect + train
package_d5.sh # package_d5 collect + train
bread_lever_bowl.sh # bread_lever_bowl collect + train
ringpeg.sh # ringpeg collect + train
yp_train_only.sh # train-only (after manual merge)
package_d5_train.sh # package_d5 train-only
jobs/ # Slurm sbatch wrappers
merge_shards.py # merge N shard dirs into one output dir
upload_results.sh # HuggingFace upload script
1028_run/ # Speedup run (4x parallel collection + AMP training)
tasks/
yp_task.sh # battery / ring / plug — 4x parallel
yp_task_1031.sh # plug1ramp / battery2 / eraser1 — 4x parallel (2026-10-31)
package_d5.sh
bread_lever_bowl.sh
ringpeg.sh
jobs/ # Slurm sbatch wrappers
merge_shards.py
```
Code lives at:
```
/lustre/fsw/portfolios/nvr/users/jtremblay/yu/1027_sim_tasks/share_v5_sim2real_20260929/code/
ur5_lab_code/ # DP training (Hydra/PyTorch)
ur5_import/ # sim collect utils, build_scene, pack_zarr
sim_data/ # symlinks to per-task data dirs
train_runs/ # training output dirs (checkpoints)
```
---
## Environment
Two conda envs on Lustre (`/lustre/fsw/portfolios/nvr/users/jtremblay/yu/miniconda3/`):
| Env | Purpose |
|-----|---------|
| `simfoundry` | Isaac Sim 5 + OmniGibson (data collection) |
| `ur5dp` | PyTorch + Diffusion Policy (training) |
Container: `robolab-isaacsim5.sqsh` (Isaac Sim 5, CUDA 12.8).
### Key env vars for collection
```bash
OMNIGIBSON_HEADLESS=1
OMNI_KIT_ACCEPT_EULA=YES
OMNIGIBSON_RENDER_VIEWER_CAMERA=0
OMNIGIBSON_DATA_PATH=.../BEHAVIOR-1K/datasets
OMNIGIBSON_GPU_ID=<0|1|2|3> # physical GPU index (do NOT set CUDA_VISIBLE_DEVICES)
TORCH_COMPILE_DISABLE=1
TORCHDYNAMO_DISABLE=1
```
---
## Data Collection
### Camera quality
Cameras render at **native resolution** from `ur5_import/cfg/scene_real.yaml`:
- `cam0` (RealSense D455): **1280×720**, 10 fps
- `cam1` (FLIR Blackfly): **1440×1080**, 10 fps
Videos encoded with `libx264 yuv420p` via imageio/ffmpeg. Training crops both to **224×224**.
4x parallel collection does **not** affect render quality: each process gets its own Isaac Sim instance on a dedicated physical GPU (`OMNIGIBSON_GPU_ID=0/1/2/3`). Vulkan renderer + PhysX are both pinned to the same physical GPU.
### Per-task collect scripts
All tasks follow the same pattern: N parallel Isaac Sim processes → merge shards → pack zarr → train.
```
sbatch 1028_run/jobs/battery_job.sh
sbatch 1028_run/jobs/ring_job.sh
sbatch 1028_run/jobs/plug_job.sh
sbatch 1028_run/jobs/package_d5_job.sh
sbatch 1028_run/jobs/bread_lever_bowl_job.sh
sbatch 1028_run/jobs/ringpeg_job.sh
sbatch 1028_run/jobs/plug1ramp_job.sh # 2026-10-31 tasks, see below
sbatch 1028_run/jobs/battery2_job.sh
sbatch 1028_run/jobs/eraser1_job.sh
```
### 2026-10-31 tasks (`yp_task_1031.sh`)
Same pipeline and settings as `yp_task.sh` (4 shards on GPU 0-3, seeds 200-599, max-kept 15/15/15/20, `--dart --start-offset`, merge -> pack_zarr -> AMP training, 130 epochs). Only the collector script differs:
| `TASK=` | Script | Task | Smoke test (yu_workstation, no video) | Time per demo (with video, RTX 4090) |
|---|---|---|---|---|
| `plug1ramp` | `yp_plugrack.py --task plug1ramp` | white plug from the grey ramp -> rotate wrist 45° -> into the outlet | 12/12 | ~200 s |
| `battery2` | `yp_battery2.py` | black battery -> grey holder, then white battery -> white holder | 12/12 | ~320 s |
| `eraser1` | `yp_eraser.py` | pick the eraser, wipe two black marker lines and fine stains off the fixed whiteboard (3 lanes), put the eraser back | 12/12 | ~350 s |
Setup (once): extract `1027_sim_tasks/tasks/yu_parts_tasks_1031.tar.gz` into `$TASKS` (= `.../yu/1027_sim_tasks`). It only adds files to `yu_parts_tasks/` (the shared scripts are byte-identical to the 10-27 package) and brings the new objects in `yu_parts_tasks/object_assets/objects/` (`grey_plug_ramp_test`, `white_plug_ramp_test`, `white_board`, `blue_eraser`); `yp_task_1031.sh` links them into `yu-assets` and stops with an error if they are missing. Train-only reruns: `jobs/<task>_train_job.sh` (uses `train_only.sh`, `NAME=<task>_v1`).
The eraser1 stains are a dynamic texture on the board face material (`omni.ui.DynamicTextureProvider`); the extra npz key `ink_left` is the stain fraction left per step. Details for all three tasks: `README_1031.md` inside the package. Example episodes (seed 0, cam0 | cam1): `1027_smoke_video/1031/`.
### Shard merge
```bash
python3 merge_shards.py shard_a/ shard_b/ shard_c/ shard_d/ final_dir/ 50
```
Merges episode npz + videos, renumbers indices, writes `collect_meta.json`.
---
## Training
### Config
- Config: `train_baseline_diffusion_unet` + `task=sim_baseline`
- Policy: `DiffusionUnetBaselineAugPolicy`
- 130 epochs, `checkpoint_every=129` → saves `latest.ckpt` at end
- Cam augmentation: cam1 dropout 0.1, aug_p 0.5, erase_p 0.3, cam0 aug_p 0.3
### Speed comparison
| Setting | it/s | Batches/epoch | Time for 130 epochs |
|---------|------|---------------|---------------------|
| `1027_run`: batch=64, workers=4, FP32 | ~1.1 | 324 | ~10h |
| `1028_run`: batch=256, workers=16, AMP (FP16) | ~2.5 | 39 | **~78 min** |
**Measured speedup: 7.7× on training** (tested on battery_v1, L40s GPU).
**Speedup sources:**
1. **batch_size 64→256**: 4× fewer batches per epoch
2. **AMP mixed precision** (`+training.use_amp=True`): FP16 forward/backward via `torch.cuda.amp.GradScaler` + `autocast`, ~2× throughput on L40s tensor cores
3. **num_workers 4→16**: parallel zarr.zip cache loading
AMP is patched into `train_equi_workspace.py`. Disable with `+training.use_amp=False`.
### Collection speedup
| Setting | Total collect time (50 demos) |
|---------|-------------------------------|
| `1027_run`: 2x parallel (GPU 0+2) | ~5–8h |
| `1028_run`: 4x parallel (GPU 0/1/2/3) | **~2.5h** |
**Measured: battery_v1 50 demos in 2h 26min** (4x parallel, 15+15+15+20 shards).
~270s per demo × 50 demos / 4 GPUs = ~56 min minimum. With Isaac Sim init overhead (~3 min/proc × 4 staggered) total ~2.5h.
---
## Known Issues
### Isaac Sim GPU crash (RenderGraph Error)
After ~3–4h / ~18 demos per process, Isaac Sim hits GPU memory leak. Solution: collect in short shards (max-kept=15–20), then merge.
### toaster_lever skipped
`black_toaster_lever` USD crashes `add_reference_to_stage` at the C++ layer in Isaac Sim 5. Not fixable without USD rebuild.
### OmniGibson vision_sensor patch
`vision_sensor.py` `_remove_modality_from_backend` raises `TypeError` during multi-episode teardown. Fix: wrap `detach()` in `try/except (TypeError, AttributeError)`.
### OMNIGIBSON_GPU_ID vs CUDA_VISIBLE_DEVICES
Isaac Sim uses `OMNIGIBSON_GPU_ID` for both Vulkan rendering (`active_gpu`) and PhysX CUDA (`physics_gpu`). **Do not** combine with `CUDA_VISIBLE_DEVICES` — it causes PhysX to request an invalid device ordinal.
---
## HuggingFace Upload
Dataset: `yqi19/dagger`
```
1027_sim_demo/<task>/collect_meta.json
1027_sim_demo/<task>/replay_buffer.zarr/
1027_sim_ckpt/<task>/latest.ckpt
```
Upload script: `1027_run/upload_results.sh`
---
## Cluster
- Account: `nvr_lpr_rvp`, QOS: `normal`, Partition: `batch_long`
- Node spec: L40s × 4 GPUs per node, 56 CPUs, 48GB VRAM per GPU
- `--comment=fact_off` on all jobs
- Container mounts: `/lustre/fsw/portfolios/nvr/users/jtremblay:/lustre/fsw/portfolios/nvr/users/jtremblay`
|