79 robot datasets, graded one at a time

We graded open-gigaai/GigaBrain-0.7-SampleData, the 1,372-episode release your team put out with the GigaBrain-0.7 report. In each push_button subset 6 of the 25 action dimensions never move across any frame, three of them holding exactly one value, and 105 of the 360 push button episodes hold the entire right arm still from first frame…

We went through RoboArena DataDump 02-03-2026, one of the fourteen sources FailBench selects and grades under the labels its own recorders assigned. Scoring every rollout by its uncredited seconds, meaning recorded time times one minus partial_success, 1,237 of the 9,589 leave at least a full 399 step cap uncredited, and 534 of those are…

We graded vegebirrd/SynGrasp-1B, the corpus your GraspVLA paper curates and the PKU-EPIC README links to. In 17,540 episodes read from 12 of the 1,256 shards, 583 carry episode.success false and ship anyway, and they concentrate hard, pick_up_propeller loses 17 of 63 and pick_up_tennis_racket 14 of 53 while 92 of the 294 object classes n…

We graded your Community Curated v0, all 981 sub-datasets and 52,507 episodes and 229.5 hours of it. Giving every judged episode one number, the share of its judge-labelled frames where the gripper and the manipulated object are both in shot, 1,365 episodes come in below half and 160 sit at zero, meaning the camera never shows the work a…

We graded HRI CUES, the Zenodo release you are a named creator on, recorded on a Furhat robot. Every one of its 590 exchanges carries three expert enjoyment scores, and only 147 of them are unanimous, while on 112, 19.0%, the panel splits by two or more points on a five point scale. Training the scorer without those 112 beat training on…

We ran your team's CaliBench release through Gantry, all 1,728 committed rollouts across the 54 scene and model cells. Scoring each cell by how many of its 32 generations carry a legible outcome, 16 cells come out under half and those 16 hold 447 of the release's 611 illegible generations, so dropping 29.6% of the corpus still keeps 1,05…

Michael's post on the robot data problem names InternData-A1, so we graded it, all 382 ARX Lift-2 subsets and 236,191 episodes of the LeRobot v3 build. 2,328 episodes carry an action record that is bit identical to another episode sitting next to it in the same subset, and the three folding subsets are most of it, 1,708 episodes holding…

We went through Gongsta/trlc_tshirt_folding, the dataset your TRLC Data Explorer lists first, recorded on DK1 arms by a customer rather than published by you. Four of its 532 takes ship with a commanded gripper that never opens or closes for the whole recording, and those four are the four shortest takes in the set, the shortest 7 frames…

We went through the eleven LeRobot v2.1 datasets published under TrossenRoboticsCommunity on Hugging Face, 560 episodes and 349,886 frames on Trossen AI arms. Capture stops on a timer and not on the task, so 14.5% of all action frames are an arm commanded to hold position, 56 episodes run five seconds or more past their last commanded mo…

We graded lerobot/community_dataset_v3, the corpus your org publishes out of the LeRobot record workflow that leLab is the front end for. Across all 1,755 recordings and 72,418 episodes, 255 recordings carry the byte-identical sentence Grasp a lego block and put it in the bin, a sentence 165 separate contributors ship, while every other…

We graded viamrobotics/box-opener, the box-opening release your Hugging Face org published and fine-tuned smolvla-box-bot on. All 110 episodes carry the same task string, open box flaps, but the demonstrator speeds up as the session runs, so episode length falls 27% from the first half to the second and the 19 fastest episodes average 19…

We graded so101-pick-and-place, the pick and place set your org published on Hugging Face. 25,618 of its 52,796 frames, 48.5%, are frames where the six commanded joint angles move under half a degree in total from the frame before, because 70 of the 72 takes are locked inside a 702 to 736 frame window, so the recorder keeps rolling long…

I graded youtalk/iknet, the Open Manipulator X recording you committed to your own repo in December 2020, as a demonstration of the method rather than a comment on your current work. The last 10.4 seconds of dataset/train, 1,040 of its 33,186 pose to joint pairs, are the arm resting on the table inside two 5 mm cells with 28 distinct joi…

We went through the Argoverse 2 Sensor validation split you co-authored, all 150 logs and all 1,701,935 annotated cuboids. Cuboid size is fixed per track there, and 16,254 distinct sizes cover the 18,395 tracks, except that 1,856 tracks reuse one of just sixteen sizes and 1,427 carry the identical 4.03 by 1.74 by 1.41 metres. Those 1,856…

We graded the 20 Stretch dex-teleop takes you published as cpaxton/pickup_pink_cup, recorded off your own cpaxton/readme-updates branch in August 2024. Episode 7 runs the full reach and lift, 0.5714 down to 0.2634 and back up, with the gripper command held at its fully open extreme on all 125 frames, so a pick that grasped nothing sits i…

We graded your team's InstructMove release on Hugging Face, the whole 500 episode set. In all 300 PiperX dual arm episodes one of the two arms never moves, 7 of the 14 action dimensions hold at most two values from an episode's first frame to its last, and which arm is still flips from episode to episode, 160 left and 140 right. We dropp…

We went through CFG-Bench, the benchmark your team released and you are a corresponding author on, and gave each of its 9,615 multiple choice items one number, how many of three text-only pickers land on its correct option. Scoring the options by word overlap with your team's released caption for that same video is right 47.2% of the tim…

Your UniT and DIAL repos both open data prep by downloading nvidia/PhysicalAI-Robotics-GR00T-Teleop-Sim, so we graded that release, all 24 task directories, 24,000 takes, 5,820,277 frames, 80.84 hours at 20 fps. Every one of the 24 directories declares 1,000 episodes and ships a declared train split that stops at episode 100, the same st…

We went through the t4-sample release attached to tier4_perception_dataset, the repo you maintain, and read num_lidar_pts on every annotated 3D box in it. 151 of the 658 boxes hold no lidar return at all, and range does not explain it, the median removed box sits 36.8 m from the lidar against 70.6 m for the ones kept. 44 of the 96 boxes…

We graded Menlo/robot-hand-poses-train, the corpus your Poseless-3B model card names as its training set. Reading 11,000 of the 100,000 poses, 20.8% of them ask one of the four tendon coupled finger pairs to hold its two joints more than 1.20 rad apart, which one actuator pulling both cannot do, and the worst asks for 1.5638 rad. Two of…

We graded gbionics/cmu-fbx, the CMU FBX set your team published and the only public dataset in the gBionics org. Scoring all 2,548 rows of metadata.csv by how many clips a query for that exact label returns once case is normalised, 201 carry the literal string Unknown, 98 carry only the word clean, 94 only the word walk, and 5 rows point…

We went through UAVScenes, the ICCV 2025 release you co-authored with Sijie Wang and the rest of the lab, all 20 flights and all 24,126 hand-labeled camera frames. 1,305 of them, 5.4%, were shot inside a stretch of 3 seconds or longer where the RTK stream has the drone under 0.1 m per second, and a further 386 carry a semantic mask that…

We graded FreeTacMan on OpenDriveLab's Hugging Face org, all 6,428 episodes across the 46 task folders. Every episode reduces to one number, the largest absolute TCP position its own trajectory file writes, and that number splits the release in two with nothing in between, 5,629 episodes run from 197.92 to 542.91 and 799 run from 0.271 t…

We went through WildWorld part 1, the 3,434 episode release your lab put on Hugging Face, working only from the public file tree. 349 of those episodes spend under 4.54 Mbps on their RGB stream against a 7.57 Mbps median at the same 720p and 30 fps, and 133 of them also land in the lowest tenth of the corpus on the lossless depth stream,…

We graded the SUOP release your lab published. Object class is recoverable from the byte size of point_cloud.xyz alone for 47.3% of the 1,554 recordings against a 20.5% chance rate, and once range is held fixed that climbs to 69.8% inside the 10 m set, so a model can score by counting points instead of reading shape. We also found 130 re…

We graded lerobot/pusht, the dataset your lerobot-doctor README runs its worked example on and the default env in lerobot-bench. Your short-episode check reports 2 of 10 on the sample in that README, and at full scale it is 53 of the 206 episodes under the 100-step action chunk ACT uses by default, 4,319 of 25,650 frames, with next.succe…

We graded rerun/v30_apple_storage, the three episode LeRobot v3 fixture in Rerun's Hugging Face org, cut from pollen-robotics/apple_storage. Its meta/info.json declares 14,983 frames and a train split of 0 to 50, which are the parent release's own numbers, while the repo holds 3 episodes and 899 frames, so 14,084 declared frames are not…

We graded murobotics/tblock-all-piper-clean-v5-bi_piper_follower, the HandUMI training set your murobotics org published on 5 September. The left gripper action holds a single value from the first frame to the last in 90 of its 176 episodes, 21,820 of 47,539 frames, while the right gripper moves in every episode and never ranges under 28…

I graded hmunn/rapt-g1-ood, the release your RAPT README links as its dataset, by reading every episode's length straight out of the npz archive index rather than downloading the arrays. On the 50 run real robot split, log duration by itself separates the 39 anomalous runs from the 11 nominal walks at 0.993 AUROC, and a threshold set at…

We graded physical-intelligence/aloha_pen_uncap_diverse, the repo openpi hard-codes in its pi0_aloha_pen_uncap config, which your org hosts as a LeRobot conversion of a BiPlay subset recorded at Berkeley. Every one of the 123 episodes is exactly 300 frames, a uniform 6 second window at 50 fps, and the window overruns the demo often enoug…

We went through the PlayWorld human demonstration release on Tenny Yin's Hugging Face account, the account your lab's open-world repo tells people to download from. Every one of the 925 episodes carries success 1, in all six version and split directories, so that field cannot separate a good demonstration from a bad one, and 102 of them…

We graded robocasa-PnPCounterToSink-MG_100_F_472, a RoboCasa upload owned by DAVIAN Robotics, the group you founded and are still listed in on Hugging Face, not anything from Holiday Robotics. Every episode in it that carries a success label is labeled failure, 472 of 572, and all 472 are exactly 700 frames long, so the label is readable…

All twelve arm joints read 0.0 for every one of the 3,037 frames of episode 3, which still carries next.success True, while six cameras record the task being completed.

561 of 663 VLA action segments span zero time, confirmed twice from start and end frames and from the release's own duration column.

All 176 marker tracking files name a D05 video, and no D05 video exists anywhere in the release. 22 recordings have every marker frozen.

Six topics are declared in all 1,639 recordings and carry zero messages in every one, including robot_pose and waypoints.

699 of 1,275 episodes hold the whole left arm at one value and 544 hold the right, and the two sets are disjoint.

left_gripper_width is constant for the whole episode in 56 of 122, while the right gripper actuates in every one.

Two episode files carry no action column at all, and the data is 532 rows short of what meta declares.

Laid out as LeRobot with no action column anywhere, and the task strings are the integers 0 to 3.

A 5,380 frame block carries an episode index that already belongs to a different recording, so 217 of the 453 episodes load frames that are not theirs, with the index, frame number and length all still matching the metadata. Reproduced against lerobot 0.6.1.