Datasets:

cagataydev
/

strands-isaaclab-allegro-reorient

Partially trained: 1500 of the default 5000 iterations; success rate 0.55 and still improving. See Limitations. The PPO policy was trained with strands' isaaclab train_policy provider (PR #4227); its rollouts, written with strands' DatasetRecorder, are this LeRobot v3 dataset.

4 recorded envs (2×2) from the strands camera — mp4. Policy: cagataydev/strands-isaaclab-allegro-reorient-policy.


episodes / frames	12 / 3351 at 30 fps (up to 300 frames = 10 s @ 30 fps; 11 full, 1 drop)
camera	`observation.images.front` 320×240 RTX, fixed (eye 0.35,-0.5,0.8 → target 0,-0.17,0.52)
`observation.state`	147-D = 16 joint pos + 124-D policy observation (`policy_obs.*`) + root pos (3) + root quat xyzw (4)
`action`	16-D raw policy action: joint position targets for the 16 hand joints, 30 Hz
task string	"reorient the cube in hand to match the goal orientation"
episode return (mean)	23.9 (range 1.2 – 42.5)
checks	strands `verify_dataset` ok · `LeRobotDataset` load ok · video decode ok · NaN/inf = 0 · Hub round-trip ok

Results (the policy that generated this data)

Trained with 8192 parallel envs × 1500 PPO iterations (task default: 5000) (295 M env steps) in 61 min 42 s on one NVIDIA L40S (Newton/MJWarp). Mean reward -0.25 → 25.1 (best) → 21.9 (last); reorientation success 0.55 (still rising when stopped); cube orientation error ≈ 1.2 rad, position error ≈ 3.8 cm; throughput median 112 k env-steps/s (max 114 k, GPU shared with other runs part of the time).

PPO iteration	mean reward	reorientation success	orientation error (rad)	env-steps/s
0	-0.25	0.003	2.11	47,900
250	4.81	0.012	1.88	112,600
500	8.89	0.109	1.17	78,841
750	12.00	0.245	1.08	112,692
1000	15.78	0.389	1.12	113,070
1250	21.93	0.488	1.14	112,959
1499	21.90	0.549	1.20	113,334

How it was made with strands-robots

# Isaac Lab in its OWN venv (its pins clash with strands; strands never imports it)
uv venv --python 3.12 ~/il && uv pip install --python ~/il/bin/python --prerelease=allow \
  --index https://pypi.nvidia.com --index-strategy unsafe-best-match "isaaclab[rsl-rl,isaacsim]==3.0.0rc1"
export ISAACLAB_PYTHON=~/il/bin/python
export OMNI_KIT_ACCEPT_EULA=YES          # you accept the NVIDIA Omniverse / Isaac Sim EULA yourself
pip install "git+https://github.com/cagataycali/robots@feat/isaaclab-trainer"   # strands-robots with PR #4227

1. Train (strands `train_policy`, `isaaclab` provider)

As an agent tool call (the train_policy tool is a Strands @tool):

from strands import Agent
from strands_robots.tools.train_policy import train_policy

agent = Agent(tools=[train_policy])
agent("Train the Allegro hand to reorient a cube in hand with the isaaclab provider: task Isaac-Reorient-Cube-Allegro, 1500 iterations, seed 1.")
# -> train_policy(action="train", provider="isaaclab", steps=1500, seed=1, output_dir="runs/c2_allegro_reorient",
#                 extra={"task": "Isaac-Reorient-Cube-Allegro", "timeout_s": 10800})

As plain Python (exactly what produced this run):

from strands_robots.tools.train_policy import train_policy

job = train_policy(action="train", provider="isaaclab", steps=1500, seed=1, output_dir="runs/c2_allegro_reorient",
                   extra={"task": "Isaac-Reorient-Cube-Allegro", "timeout_s": 10800})   # task default: 8192 envs, Newton/MJWarp; default 5000 it — we ran 1500
train_policy(action="status", provider="isaaclab", job_id="<job_id from the result>")

Under the hood: python -m isaaclab train --rl_library rsl_rl --task Isaac-Reorient-Cube-Allegro --max_iterations 1500 --seed 1 in $ISAACLAB_PYTHON. Job id of this run: isaaclab-20260929-090826-dc64d8c88224. Docs: docs/learn/training/isaaclab.md · PR: strands-labs/robots#4227.

2. Record (strands `Policy` + `DatasetRecorder`)

The final checkpoint was rolled out and recorded with examples/record_trained_policy.py (included): rebuild the task env in play mode with an RTX camera → load model_1499.pt with rsl_rl and export TorchScript/ONNX → wrap the actor as a strands Policy (RslRlJitPolicy, max |Δa| vs rsl_rl = 8.3e-07) → step with policy.get_actions_sync(...) → write every frame through strands DatasetRecorder (LeRobot v3) → verify with strands verify_dataset.

OMNI_KIT_ACCEPT_EULA=YES PYTHONPATH=/path/to/strands-robots $ISAACLAB_PYTHON examples/record_trained_policy.py \
  --task Isaac-Reorient-Cube-Allegro --checkpoint model_1499.pt --episodes 12 --frames 300 \
  --cam fixed --cam_name front --eye 0.35,-0.5,0.8 --target 0,-0.17,0.52 \
  --task_str "reorient the cube in hand to match the goal orientation" --robot_type allegro_hand --root out/ds --repo_id cagataydev/strands-isaaclab-allegro-reorient
$ISAACLAB_PYTHON examples/record_trained_policy.py --verify out/ds --repo_id cagataydev/strands-isaaclab-allegro-reorient

Use it

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("cagataydev/strands-isaaclab-allegro-reorient")
print(ds.num_episodes, ds.num_frames)            # 12 3351
x = ds[0]; print(x["observation.state"].shape, x["action"].shape, x["observation.images.front"].shape)
# (147,) (16,) (3, 240, 320)

Train an imitation policy on it with strands:

from strands_robots.tools.train_policy import train_policy
train_policy(action="train", provider="lerobot", dataset_repo_id="cagataydev/strands-isaaclab-allegro-reorient",
             output_dir="runs/act_allegro", steps=20000, batch_size=32, extra={"policy_type": "act"})

Provenance

strands-robots: feat/isaaclab-trainer @ fa66fc68 — strands-labs/robots#4227 (isaaclab train_policy provider; DatasetRecorder; verify_dataset)
Isaac Lab 3.0.0rc1 · Isaac Sim 6.1.0.0 · Newton / MJWarp (task default physics) · rsl-rl-lib 5.4.1 (PPO) · lerobot 0.6.1
GPU: 1× NVIDIA L40S (46 GB), shared with the cable-lift and Ant scaling runs for part of training
Seeds: training seed 1 (params/agent.yaml, params/env.yaml); recording seed 7
Training job: isaaclab-20260929-090826-dc64d8c88224, 2026-09-29

Limitations

Simulation only; no real Allegro hand was used; no sim-to-real claims.
Release candidates: Isaac Lab 3.0.0rc1 / Isaac Sim 6.1.0.0. Trained on the task's default Newton/MJWarp physics; the checkpoint does not remember the preset (IL-X-011) — replay on the same physics.
Partially trained: 1500 of the task's default 5000 PPO iterations (stopped for GPU time). Success rate 0.55, cube position error ≈ 3.8 cm, orientation error ≈ 1.2 rad at the last iteration, and still improving when stopped. Expect a clearly better policy with the full 5000 iterations; compare with the fully trained Shadow-hand reorientation (0.93).
Recording: 12 parallel envs, 300 frames (5 s @ 60 fps) each from one fixed camera; 11 of 12 episodes ran the full 10 s time-out, 1 ended early (cube dropped, 51 frames). observation.state = 16 joint pos + 124-D policy observation + root pos/quat (147-D), not a standard LeRobot robot state.
create_policy("rl") in strands cannot load rsl_rl checkpoints yet (IL-X-006); use the exported TorchScript + the wrapper above.

License

Card choice: license: other — our generated data / weights under CC-BY-4.0, plus NVIDIA notices. Why:

Trajectories, rendered camera video, playback clips and trained weights are user-generated content made with NVIDIA Isaac Sim / Isaac Lab; the NVIDIA Omniverse License Agreement (governs Isaac Sim 6.1, isaacsim/LICENSE.txt) §2.1 allows distributing "user generated content that you develop using Omniverse, such as video, audio, stills, models, 3D assets and screen captures". We release it under CC-BY-4.0.
No NVIDIA Content is redistributed: the Allegro Hand USD (Robots/WonikRobotics/AllegroHand/allegro_hand_instanceable.usd) and the DexCube USD (Props/Blocks/DexCube/dex_cube_instanceable.usd) come from the Isaac Lab / Isaac Sim asset packs and are not in this repo (params/env.yaml only references their paths). "Allegro Hand" is a product of Wonik Robotics; no endorsement by Wonik Robotics or NVIDIA is implied.
params/*.yaml are Isaac Lab configurations (BSD-3-Clause); examples/* are Apache-2.0 like strands-robots. Running Isaac Sim requires your own acceptance of the NVIDIA Isaac Sim / Omniverse EULA. Full text: LICENSE.md.