---
license: mit
library_name: lerobot
pipeline_tag: robotics
tags:
- robotics
- imitation-learning
- diffusion-policy
- lerobot
- pusht
- sample-efficiency
---

# Diffusion Policy · PushT · sample-efficiency study

Four **Diffusion Policy** checkpoints for the **PushT** manipulation task (LeRobot / `gym-pusht`),
from a controlled study: *under a small budget of demonstrations, does aggressive regularization
buy sample-efficiency for imitation-learning policies?*

Trained & evaluated on **AMD ROCm** (Radeon AI PRO R9700, gfx1201).

## Checkpoints

Each folder holds the final 150k-step policy (`model.safetensors` + config + pre/post-processors),
loadable with LeRobot.

| Folder | Recipe | Demos | Success (pc_success, 100 eval eps) |
|---|---|---:|---:|
| `standard-200demos` | standard | 200 | **29%** |
| `enhanced-200demos` | enhanced | 200 | **29%** |
| `standard-100demos` | standard | 100 | **8%** |
| `enhanced-100demos` | enhanced | 100 | **19%** |

- **standard** = LeRobot defaults.
- **enhanced** = aggressive regularization: weight decay `1e-3` + image-transform data augmentation.

## Result

| Demos | Standard | Enhanced | Δ |
|---:|:---:|:---:|:---:|
| 200 (data-rich) | 29% | 29% | 0 pts |
| 100 (data-scarce) | 8% | **19%** | **+11 pts** |

**Preliminary evidence** suggests that, under a 100-demo budget, enhanced regularization improves PushT
success from **8% to 19% (+11 pts)**, while showing **no gain at 200 demos** (29% = 29%). This is consistent
with regularization helping most when demonstrations are scarce (inspired by Konwoo et al. on data-constrained
pre-training). Single seed — a strong signal, not yet a settled claim.

## Training

- Policy: Diffusion Policy · Task: PushT (sim)
- **150,000 steps, seed 0**, eval on 100 episodes
- Hardware: AMD Radeon AI PRO R9700 (ROCm), ~11 steps/s
- `results/` contains the raw eval CSVs (`sweep_b*.csv`).

## Limitations

- **Single seed (0)** — deltas are a signal, not yet a claim; multiple seeds + error bars pending.
- **Baseline below the published reference (~65%)** — the remaining gap is batch size (8 vs ~64);
  addressable with gradient accumulation.
- Preliminary: enough to show the effect, not yet to publish.

## Links

- Code & write-up: https://github.com/enyolanev-bit/sample-efficient-imitation
- Case study: https://enyolanev-bit.github.io/sample-efficient-imitation/
- Interactive results: https://enyolanev-bit-sample-efficient-imitation.static.hf.space