README.md

metadata

license: mit
library_name: lerobot
tags:
  - lerobot
  - envhub
  - robotics
  - simulation
  - benchmark
  - ladder
  - franka
  - droid
  - isaac-sim
  - isaac-lab

LADDER RoboLab-Long

44 tasks of 1 to 10 subgoals for a Franka arm with a Robotiq gripper (the DROID setup), built on RoboLab and Isaac Sim 6 / Isaac Lab 3. It is one of the three suites of LADDER, a long-horizon manipulation benchmark, packaged as a LeRobot EnvHub environment. Each simulator needs its own Python environment, so each suite is its own repository:

Suite	Repository	Simulator	Robot	Tasks
RoboLab-Long	this one	Isaac Sim 6	Franka (DROID)	44
SimplerEnv-Long	INSAIT-Institute/LADDER-SimplerEnv-Long	SAPIEN 3 / ManiSkill 3	WidowX	54
RobotArena-∞-Long	INSAIT-Institute/LADDER-RobotArena-Long	Genesis	WidowX	47

Evaluate a policy

From a LeRobot checkout, with Docker, the NVIDIA Container Toolkit and an RTX GPU:

pip install -e ".[dataset,evaluation,remote]"  # plus your policy's extra
lerobot-env-server --envhub=INSAIT-Institute/LADDER-RoboLab-Long@v1.0.0 --docker=true
lerobot-eval --policy.path=<your DROID checkpoint> \
  --env.type=envhub --env.hub_path=INSAIT-Institute/LADDER-RoboLab-Long@v1.0.0 \
  --env.server.endpoint=tcp/127.0.0.1:7448 --env.task=comp_bin_2 \
  --eval.batch_size=8 --eval.n_episodes=8 --eval.use_async_envs=false

• The first --docker=true builds the simulator image (about 29 GB, 10 to 20 minutes); later starts reuse it.
• --env.task takes a task, a comma-separated list, or nothing for the whole suite (task list). A task's 8 episodes run as one batch: keep the batch size at 8 on a 48 GB GPU, 2 on a 24 GB one.
• It runs only behind the server, like LeRobot's own RoboLab integration: Isaac Sim needs NumPy 2.3, which LeRobot does not accept. The server starts Isaac Sim before it reports ready (about 5 minutes the first time, compiling shaders).

Interface

Observation	pixels.main (over-the-shoulder) and pixels.wrist, 224x224 RGB; agent_pos: 7 joint positions and gripper (0 open, 1 closed)
Action	7 absolute joint positions and gripper (above 0.5 closes), as LADDER runs pi0.5 (pi05_droid_jointpos)
Control rate	15 Hz
Score	pc_success: LADDER's task success. avg_sum_reward: its subgoal progress at the end of the episode (also info["metrics"]["subgoal_progress"])

Relation to LADDER

ladder/ is LADDER's code at release v1.0.0, unchanged, run with the settings of LADDER's own RoboLab server; requirements.txt is LADDER's lock of that environment. Isaac Lab, RoboLab and the assets (about 1 GB) are downloaded at the versions LADDER pins on first use. Its reset frames match LADDER's recorded episodes, and pi0.5 run through it succeeds about as often as in LADDER's published episodes (19 and 16 of 32 against 21 on four tasks, within run-to-run variation). LADDER's own harness stays the reference for the leaderboard.

Citation

@article{ladder2026,
  title   = {{LADDER}: A Long-Horizon Robot Manipulation Benchmark},
  author  = {Petkov, Petko and Albanese, Giuliano and Nikolov, Nikolay and Van Gool, Luc and Tremblay, Jonathan and Paudel, Danda Pani},
  journal = {arXiv preprint arXiv:2610.18053},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.18053}
}

License

MIT, as LADDER. RoboLab is Apache-2.0 and its assets mostly CC BY-NC-SA 4.0 (THIRD_PARTY_NOTICES.md). Running it accepts the NVIDIA Isaac Sim license (env.py sets OMNI_KIT_ACCEPT_EULA=YES).