rskill-act-aloha-aloha_insertion-fp32

OpenRAL rSkill (custom example) — ACT (Action Chunking Transformer) finetuned on the ALOHA bimanual peg-insertion task, packaged for OpenRAL.

This package wraps lerobot/act_aloha_sim_insertion_human with a rskill.yaml manifest that adds capability checking, license surfacing, latency budgets, and local registry integration. It does not copy model weights.

It is the harder sibling of rskill-act-aloha-aloha_transfer_cube-fp32 (cube transfer): one packaging format covering multiple task-specific checkpoints from the same paper. The runnable demo is scenes/benchmark/aloha_insertion.yaml, wired into the top-level just sim-custom recipe.

Upstream model

Field	Value
Source repo	lerobot/act_aloha_sim_insertion_human
Architecture	Action Chunking Transformer (~52M params, chunk=100)
Task	gym-aloha AlohaInsertion-v0 (bimanual peg-in-socket)
License	MIT
Paper	Zhao et al., 2023 — Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (arXiv 2304.13705)

Evaluation

Locally reproduced on the ALOHA bimanual peg insertion benchmark via openral benchmark run --suite aloha --rskill act-aloha-insertion (the aloha suite skips the transfer-cube task via evaluated_tasks) (reproduced_locally: true, see eval/aloha_insertion.json).

Suite	Episodes	Protocol	Result
aloha_insertion	50	seeds 0–49, success_key=is_success, max_steps=400	10/50 success (20%)

The paper's sim ALOHA insertion number is markedly lower than its cube-transfer figure (harder task; the upstream protocol uses different camera intrinsics); this reproduction shows the same gap, not a regression.

Supported robots

Robot	Embodiment tag	Status	Notes
ALOHA bimanual (Trossen) — gym-aloha MuJoCo	aloha	✓ sim	14-DoF (2 × 7-DoF arms with parallel grippers); MuJoCo AlohaInsertion-v0.

Same physical embodiment as the act-aloha sibling; only the task contact dynamics differ (peg insertion is harder than cube pick-and-place).

Sensors required

Key	Modality	Resolution	Format
observation.images.top	RGB camera	640 × 480	float32
observation.state	proprioception	(14,)	float32 (2 × 7-DoF joint positions)

Single top-down RGB stream, like the cube-transfer sibling; no wrist or third-person views.

Manifest summary

Field	Value
name	OpenRAL/rskill-act-aloha-aloha_insertion-fp32
version	0.1.0
license	mit
role	s1
embodiment_tags	aloha
runtime / quantization.dtype	pytorch / fp32
weights_uri	hf://lerobot/act_aloha_sim_insertion_human
chunk_size	100
is_commercial_use_allowed (derived from license)	true

Full schema: openral_core.schemas.RSkillManifest.

Run it

just sim-custom

…a one-episode BenchmarkScene demo, equivalent to:

MUJOCO_GL=egl uv run --all-packages --group sim openral benchmark scene \
  --config scenes/benchmark/aloha_insertion.yaml \
  --rskill rskill://rskills/act-aloha-insertion \
  --no-update-manifest --n-episodes 1 --save-dir example_videos

License

OpenRAL wrapper files follow the project Apache-2.0 license. The upstream weights are MIT, so the manifest declares license: mit and commercial use is allowed (is_commercial_use_allowed is true).