strands-isaaclab-reach-franka-mujoco-sim2sim

Sim-to-sim: a PPO policy trained in NVIDIA Isaac Lab through strands-robots' isaaclab train_policy provider (PR #4227), deployed unchanged on strands-robots' MuJoCo backend and recorded with strands start_recording. One strands API on both ends: train_policy(provider="isaaclab") to train, create_simulation("mujoco") + run_policy + start_recording to deploy and record.

strands MuJoCo rollouts — mp4. Policy: cagataydev/strands-isaaclab-reach-franka-policy · Isaac Lab counterpart: cagataydev/strands-isaaclab-reach-franka.

Result — the sim-to-sim gap

	Isaac Lab (training sim, Newton/MJWarp)	strands MuJoCo (this dataset)
reached pose (< 5 cm and < 0.2 rad)	11 / 12 recorded (0.983 at training)	10 / 12
episode	terminates at success	fixed 4 s, a new random command each, keeps acting after success
final position error	– (terminated)	mean 10.6 cm (drift after reaching; the 2 misses at 30 – 34 cm)

Data

• 12 episodes × 120 frames = 1440 frames at 30 fps; observation.state = 9 panda joint positions, action = 8 actuator commands (actuator1..7 position targets = default + 0.5·a, actuator8 = gripper open), observation.images.front 320×240
• task string: "reach the commanded end-effector pose"
• checks: strands verify_dataset ok · LeRobotDataset load ok · video decode ok · NaN 0 · Hub round-trip ok

How it was made with strands-robots

1. Train in Isaac Lab by a Strands Agent via train_policy(action="train", provider="isaaclab", steps=300, seed=1, learning_rate=1e-3, extra={"task": "Isaac-Reach-Franka", "num_envs": 4096}) — setup, sweep and results on the policy card. Docs: docs/learn/training/isaaclab.md.
2. Deploy + record on MuJoCo with examples/sim2sim_mujoco.py (strands venv, no Isaac):

from strands_robots.simulation.factory import create_simulation

sim = create_simulation("mujoco")
sim.create_world(timestep=1.0 / 120.0)
sim.add_robot("arm", data_config="panda")                       # MuJoCo Menagerie franka_emika_panda
sim.add_camera("front", position=[1.6, 0.9, 0.9], target=[0.4, 0.0, 0.3], width=320, height=240)
sim.start_recording(repo_id="cagataydev/strands-isaaclab-reach-franka-mujoco-sim2sim", root="ds", fps=30,
                    task="reach the commanded end-effector pose", cameras=["front"], overwrite=True)
sim.run_policy("arm", policy_object=IsaacReachOnStrands(jit, sim, seed=0),   # strands Policy adapter in the example
               instruction="reach the commanded end-effector pose", control_frequency=30.0, control_substeps=4,
               n_steps=120, fast_mode=True, n_episodes=12, reset_between=True, seed=0)
sim.stop_recording()

MUJOCO_GL=egl python examples/sim2sim_mujoco.py --jit exported/policy.pt --episodes 12 --root ds --repo_id cagataydev/strands-isaaclab-reach-franka-mujoco-sim2sim

The adapter (IsaacReachOnStrands, a strands Policy) rebuilds the 32-D Isaac Lab observation — joint_pos − default (9) | joint_vel (9) | pose command, base frame, pos + quat xyzw (7) | last action (7) — samples a command from Isaac Lab's ranges (x ∈ [0.35, 0.65], y ∈ [-0.2, 0.2], z ∈ [0.15, 0.5], pitch π, yaw ∈ [-π, π]) and starts each episode from Isaac Lab's default pose. MuJoCo dt 1/120, 4 substeps, 30 Hz control.

Use it

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("cagataydev/strands-isaaclab-reach-franka-mujoco-sim2sim")
print(ds.num_episodes, ds.num_frames, ds.fps)                    # 12 1440 30

from huggingface_hub import snapshot_download
from strands_robots.verify_dataset import verify_dataset
assert verify_dataset(snapshot_download("cagataydev/strands-isaaclab-reach-franka-mujoco-sim2sim", repo_type="dataset"), expected=12)["ok"]

Provenance

• strands-robots feat/isaaclab-trainer @ fa66fc68 (#4227) · MuJoCo 3.11.0 · lerobot 0.6.1
• policy: rsl_rl 5.4.1 PPO, 4096 envs × 300 iterations, lr 1e-3, seed 1, Isaac Lab 3.0.0rc1 / Isaac Sim 6.1.0.0, NVIDIA L40S
• rollout seed 0; sim2sim_mujoco.json in the policy repo has the per-episode errors

Limitations

• Simulation only. Different robot model than training (Menagerie MJCF vs Isaac Lab USD), different physics engine; the adapter hard-codes Isaac Lab's action scale (0.5) and default joint pose. No command resampling within an episode, so the arm drifts after reaching.
• Isaac Lab 3.0.0rc1 / Isaac Sim 6.1.0.0 are release candidates.

License

CC-BY-4.0. This dataset was rendered and simulated entirely in MuJoCo (Apache-2.0) with the MuJoCo Menagerie franka_emika_panda model (Apache-2.0); it contains no NVIDIA assets or renders. The policy that drove it was trained in NVIDIA Isaac Sim / Isaac Lab (its weights are our user-generated content, see the policy card). The example script is Apache-2.0. "Franka" is a trademark of Franka Robotics; no endorsement implied.