Timing clarification: the dashboard labels the box lift as approximately 6 seconds and the visible hold as at least approximately 15 seconds. These are estimates, not frame-verified measurements. Task completion does not establish clean success. The website’s current results and methodology contain the latest confirmed classifications.
Source notes for the current 59-second homepage demo: A documentary-style edit with the RA logo, new founder-provided narration and task-by-task findings. Times below refer to this published edit and are rounded; they are not timestamps in the original recordings.
The video combines manufacturer footage, third-party model-ranking charts and our research recordings. These are different systems, tasks and control setups—not a matched comparison or a measurement of any manufacturer’s overall reliability.
0:00–0:01.80 · Manufacturer demonstrations
The nine-panel montage labeled Unitree, Figure and 1X comes from the supplied edit. The opening plays at normal speed, with the question appearing over the NEO shot; this promotional sequence is not used for timing or scoring. Exact original upload links, dates and source offsets for individual panels are not documented. These are manufacturer demonstrations, not independently scored RoboAnalytica trials.
0:01.80–0:04.50
The existing living-room shot is shown for 2.7 seconds, using 36.50–39.20 of the supplied NEO home video at normal speed. It is manufacturer footage, not a scored trial.
0:04.50–0:09.10 · Stair assistance; 0:09.10–0:12.10 · Stair incident
The first excerpt shows outdoor stair assistance. The following full-width excerpt shows a left-hand component separating during a stair attempt. The team identifies the outdoor stair recordings as teleoperated and indoor manipulation as autonomous; Gary confirmed HANDOFF control for indoor tasks and a different locomotion/teleoperation stack for stairs.
Stairs were recorded September 11, 2025; manipulation was recorded May 15, 2026. They are separate sessions and are not pooled into a success rate. The highlighted incident is from DJI_20250911_155330_29_null_video.mp4, approximately 00:36.5–00:39.5. A component visibly separates from the left wrist/hand area around 00:37. The exact assembly is not established from video alone. This corrects our earlier blanket “no detachment” statement. It is a different attempt from DJI_20250911_154344_28_null_video.mp4 in the six-row results table, and is not included in that table’s totals.
Stair sources and recording context ↗ · Indoor excerpts and observed intervals ↗
0:12.10–0:20.50
The supplied edit displays an “LLM Intelligence” chart and an “LLM / VLA Robot Tasks” chart. The displayed numbers are not RoboAnalytica results.
The original publisher, snapshot date, chart URLs and methodology for these graphics have not yet been documented. They are third-party graphics reproduced in the edit, not verified current rankings. Their inclusion does not establish that other evaluators test only software.
0:20.50–0:25.60 · Box handling; 0:25.60–0:31.80 · Stairs; 0:31.80–0:34.00 · Shelf transfer
Our recordings illustrate hardware, software and control operating together. The mobility shot now plays for 6.2 seconds at normal speed, followed by a brief shelf excerpt. These excerpts inform the first research case study; they do not establish a completed scored benchmark, commercial reliability, or performance of every stock Unitree G1.
Recording sources, configuration limits and supporting research ↗
0:34.00–0:54.40
Three results sequences (5.6 seconds for box, 6.7 for shelf and 8.1 for stairs) illustrate recorded outcomes: an estimated 6-second lift and at least approximately 15 seconds of visible hold of a 1 kg box (partial completion); a 60-second autonomous shelf task (complete); and a 4-second teleoperated ascent of five stairs (80 cm). The displayed times describe the source actions, not the length of these illustrative excerpts. Playback remains at normal speed. The excerpts do not show every part of the measured intervals.
Indoor HANDOFF and outdoor teleoperation use different configurations. Six selected attempts include two task completions, one partial completion and three failures/interventions, with one reset. Additional unaudited recordings exist. These are retrospective research observations, not a pooled reliability rate or a passed prospective benchmark. Manipulation stage scores are box 2/3 and shelf 3/3; the completed stair run remains unscored because assistance is not established. Two other stair attempts reached partial ascent, with the hands-on score provisional. Stage scales are task-specific, not a cross-task leaderboard. Completion does not establish clean success.
0:54.40–0:59.00
The RA logo and company message close the film. The narration uses the complete founder-provided recording at its original speed, without phrase cuts or word splices. Scenes follow the narration, with a faint original electronic underscore.
This source list replaces the notes for the older, longer film. The Tesla Gen 2 balance excerpt, Figure hand-sensing ablation, dishwasher-difficulty excerpt and extended three-test-pool explanation described on that older page are not segments of this current edit.
Third-party footage belongs to its respective rights holders. Attribution does not imply endorsement or partnership.