Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
Paper • 2303.04137 • Published • 6
rskill-diffusion-pusht-pusht-fp32
OpenRAL rSkill — Diffusion Policy (Chi et al., 2023) trained on the PushT 2-D pushing benchmark, packaged for
OpenRAL.
This package wraps lerobot/diffusion_pusht
with a rskill.yaml manifest. It does not copy model weights.
Upstream model
| Field | Value |
|---|---|
| Source repo | lerobot/diffusion_pusht |
| Paper | arxiv:2303.04137 — Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (Chi et al., 2023) |
| License | Apache-2.0 |
| Parameters | ~263 M (1-D U-Net) |
| Action chunk | 8 (within horizon 16) |
| Denoising | 100 DDPM steps per chunk |
| Benchmark | PushT (gym_pusht, pymunk 2-D rigid-body) |
Per-chunk inference is dominated by the 100-step denoising loop; cached
pops are essentially free, so this is the extreme test of the
queue-drain contract in ChunkedExecutor.
Supported robots
| Robot | Embodiment tag | Status | Notes |
|---|---|---|---|
PushT 2-D pseudo-robot (gym_pusht/PushT-v0) |
pusht |
✓ sim | 2-D end-effector pushing a T block on a 512 × 512 px canvas |
Sensors required
| Key | Type | Resolution | Format |
|---|---|---|---|
observation.images.camera1 |
RGB camera | 96 × 96 | float32 |
PushT predates the multi-cam observation.images.cameraN convention: the
checkpoint's own key is observation.image. The manifest's
image_preprocessing renames slot camera1 onto it
(input_template: "observation.{cam}", aliases: {camera1: image}).
Manifest summary
| Field | Value |
|---|---|
name |
OpenRAL/rskill-diffusion-pusht-pusht-fp32 |
version |
0.1.0 |
license |
apache-2.0 |
role |
s1 |
embodiment_tags |
pusht |
runtime / quantization.dtype |
pytorch / fp32 |
weights_uri |
hf://lerobot/diffusion_pusht |
latency_budget.per_chunk_ms |
1 250 ms (warm full-chunk ≈ 1 756 ms on RTX 4070 Laptop, dominated by DDPM) |
is_commercial_use_allowed (derived from license) |
true |
Full schema: openral_core.RSkillManifest —
python/core/src/openral_core/schemas.py.
Reproduction
git clone https://github.com/OpenRAL/openral && cd openral
just bootstrap && just sync --group sim
# One-episode BenchmarkScene demo (CPU is enough):
just sim-diffusion-pusht
# which runs:
# openral benchmark scene --config scenes/benchmark/pusht.yaml \
# --rskill rskill://rskills/diffusion-pusht --no-update-manifest --n-episodes 1 --save-dir example_videos
# Full 50-episode protocol (eval/pusht.json):
openral benchmark run --suite pusht --rskill diffusion-pusht
# Sim test (gym_pusht + pymunk):
uv run pytest tests/sim/test_pusht_2d_diffusion_pusht.py -v -m sim
License
This rSkill package (rskill.yaml, README.md) is Apache-2.0 to
match the upstream weights. Commercial use is allowed
(is_commercial_use_allowed is true).
See also
robots/pusht_2d/README.md— RobotDescription manifest.scenes/benchmark/pusht.yaml— paired BenchmarkScene config (pass--rskill rskills/diffusion-pusht).docs/reference/vla_compatibility.md— VLA × Robot × Sim matrix.