---
license: apache-2.0
tags:
  - robotics
  - lerobot
  - diffusion-policy
  - imitation-learning
  - manipulation
---

# Diffusion Policy — banana-in-pot (JOINT, fp16)

Diffusion Policy trained on the **"put the right banana in the pot"** task (UR7e + GELLO
teleoperation, 2 RGB cameras), in **JOINT** action space (6 joints + gripper), using
**fp16 mixed-precision** training.

- **Checkpoint:** step 80,000 (best open-loop MAE)
- **Base library:** [LeRobot](https://github.com/huggingface/lerobot) 0.6.1 (pin `8a74e0a`)
- **Dataset:** [`Bigenlight/banana_in_pot_lerobot_v3`](https://huggingface.co/datasets/Bigenlight/banana_in_pot_lerobot_v3) — 51 episodes / 21,524 frames / 30 fps
- **fp32 sibling:** [`Bigenlight/diffusion_banana_in_pot_joint`](https://huggingface.co/Bigenlight/diffusion_banana_in_pot_joint)

## Architecture

ResNet18 (ImageNet-pretrained, trained end-to-end) + SpatialSoftmax vision encoder →
FiLM-conditioned 1D temporal U-Net denoiser (DDPM-100, epsilon prediction). ~278M params.
Images resized to 360×640, crop off. `n_obs_steps=2`, `horizon=64`, `n_action_steps=32`.

## Training

- **Precision:** fp16 via HF Accelerate `mixed_precision=fp16` (automatic GradScaler).
- Requires a `dtype` field on `DiffusionConfig` (absent upstream at this pin); launched with
  `--policy.dtype=float16`.
- Batch 8, 80k steps, AdamW, seed 1000, 45 train / 6 held-out episodes.
- **Hardware:** single RTX A4000. **~4.48 step/s, ~4.96 h** (vs ~3.49 step/s / ~6.4 h fp32 →
  ~1.25× faster, ~22% less wall-clock). No NaN/instability.

## Open-loop evaluation (DDIM-10, held-out episodes 45–50)

| | fp32 @80k | **fp16 @80k** |
|---|---|---|
| poseMAE (rad) | 0.0845 | **0.08268** |
| gripper accuracy | 0.953 | **0.960** |

**fp16 matches (marginally beats, within noise) fp32 accuracy at ~22% lower training cost.**
Select the deploy checkpoint by open-loop MAE, **not** `eval_loss` (which rises during
training for diffusion policies without indicating overfitting).

## Intended use & limitations

Research artifact. Small single-task, single-scene, real-world (noisy) dataset of 51
success-only demonstrations; offline metrics only — no closed-loop hardware success rate
measured yet. Not safety-validated for autonomous operation.