nkalyanv/argus-robot-annotations

Dense Argus Annotations for 5,000+ Hours of Open Robot Manipulation Data

Episode-level annotations for 290,026 teleoperated robot episodes (about 5,146 hours) from five public datasets, produced with the open-source Argus video-annotation pipeline and GPT-6.1 Sol (medium reasoning). Every record uses the per-episode format of Pantheon's Argus data board release: the 29 fields of a board episode, plus a source_alignment block that ties every label time to the source recording and a _release block naming the labeller. Pantheon published Argus annotations for 1,284 MolmoAct2 and 183 ABC-130k episodes; this release labels every publishable episode of those two datasets and three more.

This is an independent release, not affiliated with or endorsed by Pantheon; it uses their open-source pipeline and format. This repository contains labels only. The videos and robot data stay with their publishers (links below).

Config	Source dataset	Embodiment	Episodes	Hours	Outcome: success / failure / unclear / undone
abc130k	XDOF/ABC-130k	bimanual teleoperated arms	130,699	3,589.7	92.2 / 7.6 / 0.2 / 0.0 %
droid	lerobot/droid_1.0.1	Franka Panda + Robotiq gripper (2 external + wrist camera)	94,945	511.5	84.1 / 7.3 / 8.2 / 0.3 %
molmoact2_bimanual_yam	allenai/MolmoAct2-BimanualYAM-Dataset	bimanual YAM arms	32,223	703.8	85.4 / 13.2 / 0.4 / 1.0 %
abc_sim	amazon-far/abc sim v3 teleop data	simulated ABC bimanual rig	16,177	138.5	99.9 / 0.1 / 0.0 / 0.0 %
gm100_cobotmagic	rhos-ai/gm100-cobotmagic-lerobot	AgileX Cobot Magic (bimanual)	15,982	202.2	80.4 / 10.9 / 6.8 / 1.9 %

Each config has one train split holding every labelled episode, whatever its upstream split. Hours are the summed episode durations (duration_s). "Undone" is success_then_undone: the goal was reached and then reversed within the episode.

Loading

from datasets import load_dataset

ds = load_dataset("nkalyanv/argus-robot-annotations", "gm100_cobotmagic", split="train")
ep = ds[0]
print(ep["episode_prompt"], ep["completion"]["task_completed"])

The card declares each config's schema, so every shard loads with one consistent set of columns (tested with datasets 4.7 and 5.0). Fields a record does not have come back as null. The files are plain JSON Lines, one episode per line, and can also be read without datasets:

import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = snapshot_download("nkalyanv/argus-robot-annotations", repo_type="dataset", allow_patterns=["data/droid/*"])
episodes = [json.loads(line) for p in sorted(Path(root, "data/droid").glob("*.jsonl")) for line in open(p)]

Read this way, every line is exactly the published Argus record, with no null fill-in.

What each record contains

Field	Content
episode_prompt, instruction_variants, goal_alignment	the instruction, paraphrases, and whether the footage matches the given instruction
objects, scene_graph, state_changes	the objects involved and how their relations change over time
event_labels	per-arm timeline of actions (verb_class, object, carry_phase, progress, start t_s and end_s)
key_events	milestones (phases, placements, grasps) with outcomes
completion	task_completed (success, failure, unclear, success_then_undone), the success predicate, completion and goal times
operator_mistakes, recovery, data_issues	mistakes with category, severity, time and evidence; recoveries; recording problems
performance_review	a short prose review of the episode
arm_still_spans, idle	spans where an arm is not moving
dataset_checks	deterministic checks of the recording (frame counts, timebase, recorded jumps, gripper channels, capture QC)
camera_views, camera_labels, timesteps_s	the camera streams shown to the labeller and the instants it saw
dataset_source	the source dataset, its publisher and license
source_alignment	how to map label times to the source episode (see below)
_meta, _usage, _release	episode id, labeller, token usage and pipeline version

dataset_episode_labels (DROID only) carries the source dataset's own episode labels for comparison, copied unchanged from lerobot/droid_1.0.1; there date repeats collector_id and task_category repeats building.

Aligning labels with the source recordings

All times (t_s, end_s, *_at_s, timesteps_s) are seconds from the episode's time zero, which source_alignment.time_zero defines for each dataset:

• ABC-130k: seconds after t0_unix_ns, the capture stamp of the first camera frame of the episode's MCAP (upstream_path). For a copy resampled to its own clock with start T0 (ns) and rate fps: frame = round((t0_unix_ns + t * 1e9 - T0) / 1e9 * fps). operator_id and station come from the recording. ABC-130k is gated on Hugging Face: accept its terms before downloading.
• DROID, MolmoAct2, GM-100: seconds after the episode's first frame, so frame = round(t * fps) indexes the LeRobot frame_index of episode_index (GM-100: episode within dataset_folder). measured_fps is 15 for DROID and 30 for MolmoAct2 and GM-100. DROID and MolmoAct2 are LeRobot v3.0, where one video file holds many episodes: the episode starts at its videos/<camera>/from_timestamp (in meta/episodes/), so seek to that plus t.
• ABC-Sim: seconds after the first frame at 30 fps, indexing upstream_path's frames and state rows. Fetch a task with prepare.py --sim-data <task> from amazon-far/abc; upstream_path follows the layout it writes. The labelled tasks are sim_put_the_plastic_bottles_in_the_bin, sim_load_the_plates_into_the_dish_rack and sim_turn_the_mug_right_side_up.

How the labels were made and checked

Each episode's video streams went through the Argus pipeline (routing, labelling, board build) with GPT-6.1 Sol at medium reasoning. An episode was published only if the reply parsed and finished normally, passed Argus's schema check with no off-schema entries, had a real routing answer, a plausible measured frame rate, all deterministic checks, a timeline that covers the recording, and one of Argus's completion outcomes. An episode that failed was relabelled, with at most three attempts in all. Episodes that never passed are listed in meta/<config>/missing.tsv with reasons (ABC-130k 4, DROID 713, MolmoAct2 23, ABC-Sim 0, GM-100 0). GM-100's upstream task_00007 episode 0 has videos but no data file and was not labelled. SHA256SUMS covers every data and meta file.

Licenses

The annotations in this repository are released under CC BY 4.0. Each record's dataset_source names the license of the footage it describes: Apache-2.0 for ABC-130k, DROID, MolmoAct2 and ABC-Sim, and MIT for GM-100. Use of the underlying videos and robot data is governed by their publishers' terms.

Citation

If you use these labels, please cite this release:

@misc{nadimpalli2026argusannotations,
  title        = {Dense Argus Annotations for 5,000+ Hours of Open Robot Manipulation Data},
  author       = {Nadimpalli, Kalyan and H{\"o}ppe, Tobias and Reddy, Shashank and Bauer, Stefan},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/nkalyanv/argus-robot-annotations}},
  doi          = {10.57967/hf/11286}
}

The labels were made with Argus. Please also cite the source dataset of each config you use; the table at the top links each one.