INSAIT-Institute/LADDER-Evaluation-Episodes

LADDER evaluation episodes

📄 Paper · 💻 Code · 🌐 Website · 🤗 LADDER collection

The evaluation episodes of the methods in the main results table of LADDER: A Long-Horizon Robot Manipulation Benchmark, on every task of the three suites: 1,450 task runs and 20,240 episodes, each with its log and its video, and one successful episode of the scripted policy per task.

Layout

One directory per method, with everything about it inside: one subdirectory per suite, and one per real robot. The names are those scripts/eval.sh --method and --suite take:

<method>/<suite>/<task>/
    results.yaml                 task success, subgoal progress and failure causes over the task's episodes
    episode_log.yaml             one entry per episode (below)
    episodes/<episode>/          episode_meta.yaml, main.mp4 (RoboLab-Long: + wrist.mp4)
<method>/real_<robot>/           the method's real-robot trials, if it ran on a real robot (below)
scripted/<suite>/<task>.mp4      a successful episode of the scripted policy
real_robot.csv                   the paper's real-robot table
index.parquet                    one row per episode: suite, method, task, success, subgoals, steps, path
media/                           the GIFs and videos shown in the code repository and on the project page

suite	name	tasks	episodes per task
simplerenv_long	SimplerEnv-Long	54	24
robotarena_long	RobotArena-∞-Long	47	8
robolab_long	RoboLab-Long	44	8

method	name	low-level policy
low_level_policy	Low-level policy	Bridge policy / pi0.5
periodic_trigger	Periodic trigger	Bridge policy / pi0.5
proprioceptive_trigger	Proprioceptive trigger	Bridge policy / pi0.5
proprioceptive_verifier	Proprioceptive verifier	Bridge policy / pi0.5
volo	VoLo	Bridge policy / pi0.5
privileged_verifier	Privileged verifier	Bridge policy / pi0.5
cap_x	CaP-X	none
tiptop	TiPToP	none
psp_agent	PSP-Agent	none
gpt_policy	GPT-Policy	none

The Bridge policy drives the two WidowX suites and pi0.5 (DROID) drives RoboLab-Long. The scripted policy (scripts/eval.sh --method scripted) solves the tasks from the simulator's ground-truth object state; its videos show that every task can be solved. The dataset holds 23,905 videos (33.5 GB, H.264, at the simulator's control rate) and 1.71 GB of logs.

LADDER-Lite

LADDER-Lite is a fixed 10-task subset of the benchmark (scripts/eval.sh --task lite, 112 episodes) for testing new models quickly. Models scored on LADDER-Lite only have their own method directory, with the same layout:

method	name	VLM	tasks
gpt_policy_opus55	GPT-Policy	Claude Opus 5.5	the 10 LADDER-Lite tasks
gpt_policy_argon	GPT-Policy	Gemini 4 Argon	the 10 LADDER-Lite tasks

The tasks are relay7e and split3 (simplerenv_long), scene36_1goal, scene45_6goal, scene54_split_3 and scene76_2goal (robotarena_long), and comp_ordered_bin_4, comp_relay_3, comp_route_10 and comp_transfer_2 (robolab_long). Their rows in index.parquet carry the method name above.

Real-robot evaluations

The real-robot results of the paper, on a WidowX 250 and a Franka FR3:

real_robot.csv                                the paper's real-robot table
<method>/real_<robot>/scores.csv              the run's scores
<method>/real_<robot>/eval_config.yaml        the evaluation config
<method>/real_<robot>/episode_*/              main.mp4 (pi0.5: + wrist.mp4) and episode_meta.yaml (plan, tool calls, score)

<robot> is widowx or franka.

robot	method	low-level policy	trials
WidowX 250	low_level_policy	Bridge policy	108
WidowX 250	proprioceptive_verifier	Bridge policy	108
WidowX 250	psp_agent	none	108
Franka FR3	low_level_policy	pi0.5	39
Franka FR3	proprioceptive_verifier	pi0.5	39
Franka FR3	psp_agent	none	61

Every trial has its video. Three of the pi0.5 trials per method belong to two tasks that were not run to completion; real_robot.csv marks them complete: False, and the paper leaves them out of its comparisons.

An episode_log.yaml entry holds the instruction, success, the number of steps, the subgoals and the orchestrator's plan, the per-step subgoal progress (goal_trace), and traces of the gripper, the objects and what is held (motion_trace, object_trace, held_trace). An episode_meta.yaml holds the episode's events (plans, subgoal commands, verifications, tool calls with their reports).

Download and browse

With the benchmark's code, one suite and method (or leave out --suite / --method for all of them, and add --no-videos for the logs only), then open them in its dashboard:

scripts/dashboard.sh fetch --suite robolab_long --method psp_agent     # to ~/ladder_episodes
scripts/dashboard.sh ~/ladder_episodes/dashboard ~/ladder_episodes/dashboard_real

Or with the Hugging Face CLI alone:

hf download INSAIT-Institute/LADDER-Evaluation-Episodes --repo-type dataset --local-dir ladder_episodes --include "psp_agent/*"

index.parquet answers most questions without downloading the tree, e.g. with pandas:

import pandas as pd
df = pd.read_parquet("hf://datasets/INSAIT-Institute/LADDER-Evaluation-Episodes/index.parquet")
df.groupby(["suite", "method"]).success.mean()

What is not included

The launchers' console logs, per-run host information and GPU samples are left out: they record the machines the runs happened on, not the episodes.

License

CC BY 4.0. The RoboLab-Long videos show RoboLab assets, which are licensed under CC BY-NC-SA 4.0.

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}
}