Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
library_name: lerobot
|
| 4 |
+
pipeline_tag: robotics
|
| 5 |
+
tags:
|
| 6 |
+
- robotics
|
| 7 |
+
- imitation-learning
|
| 8 |
+
- diffusion-policy
|
| 9 |
+
- lerobot
|
| 10 |
+
- pusht
|
| 11 |
+
- sample-efficiency
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# Diffusion Policy · PushT · sample-efficiency study
|
| 15 |
+
|
| 16 |
+
Four **Diffusion Policy** checkpoints for the **PushT** manipulation task (LeRobot / `gym-pusht`),
|
| 17 |
+
from a controlled study: *under a small budget of demonstrations, does aggressive regularization
|
| 18 |
+
buy sample-efficiency for imitation-learning policies?*
|
| 19 |
+
|
| 20 |
+
Trained & evaluated on **AMD ROCm** (Radeon AI PRO R9700, gfx1201).
|
| 21 |
+
|
| 22 |
+
## Checkpoints
|
| 23 |
+
|
| 24 |
+
Each folder holds the final 150k-step policy (`model.safetensors` + config + pre/post-processors),
|
| 25 |
+
loadable with LeRobot.
|
| 26 |
+
|
| 27 |
+
| Folder | Recipe | Demos | Success (pc_success, 100 eval eps) |
|
| 28 |
+
|---|---|---:|---:|
|
| 29 |
+
| `standard-200demos` | standard | 200 | **29%** |
|
| 30 |
+
| `enhanced-200demos` | enhanced | 200 | **29%** |
|
| 31 |
+
| `standard-100demos` | standard | 100 | **8%** |
|
| 32 |
+
| `enhanced-100demos` | enhanced | 100 | **19%** |
|
| 33 |
+
|
| 34 |
+
- **standard** = LeRobot defaults.
|
| 35 |
+
- **enhanced** = aggressive regularization: weight decay `1e-3` + image-transform data augmentation.
|
| 36 |
+
|
| 37 |
+
## Result
|
| 38 |
+
|
| 39 |
+
| Demos | Standard | Enhanced | Δ |
|
| 40 |
+
|---:|:---:|:---:|:---:|
|
| 41 |
+
| 200 (data-rich) | 29% | 29% | 0 pts |
|
| 42 |
+
| 100 (data-scarce) | 8% | **19%** | **+11 pts** |
|
| 43 |
+
|
| 44 |
+
At 200 demos regularization buys nothing; at 100 demos it lifts success **8% → 19% (+11 pts, ~2.4×)**.
|
| 45 |
+
Regularization buys sample-efficiency **when demonstrations are scarce** — as hypothesized (inspired by
|
| 46 |
+
Konwoo et al. on data-constrained pre-training).
|
| 47 |
+
|
| 48 |
+
## Training
|
| 49 |
+
|
| 50 |
+
- Policy: Diffusion Policy · Task: PushT (sim)
|
| 51 |
+
- **150,000 steps, seed 0**, eval on 100 episodes
|
| 52 |
+
- Hardware: AMD Radeon AI PRO R9700 (ROCm), ~11 steps/s
|
| 53 |
+
- `results/` contains the raw eval CSVs (`sweep_b*.csv`).
|
| 54 |
+
|
| 55 |
+
## Limitations
|
| 56 |
+
|
| 57 |
+
- **Single seed (0)** — deltas are a signal, not yet a claim; multiple seeds + error bars pending.
|
| 58 |
+
- **Baseline below the published reference (~65%)** — the remaining gap is batch size (8 vs ~64);
|
| 59 |
+
addressable with gradient accumulation.
|
| 60 |
+
- Preliminary: enough to show the effect, not yet to publish.
|
| 61 |
+
|
| 62 |
+
## Links
|
| 63 |
+
|
| 64 |
+
- Code & write-up: https://github.com/enyolanev-bit/sample-efficient-imitation
|
| 65 |
+
- Case study: https://enyolanev-bit.github.io/sample-efficient-imitation/
|
| 66 |
+
- Interactive results: https://enyolanev-bit-sample-efficient-imitation.static.hf.space
|