FIELD NOTES / 001 · OCTOBER 2026
A retrospective evaluation of Unitree G1 research configurations: autonomous HANDOFF manipulation and separately controlled, teleoperated stairs. The contributing researcher confirmed the setup and selected outcomes. These observations do not rate the stock product or establish commercial reliability.
The pitch includes separately attributed manufacturer and journalist excerpts. View sources and comparison limits.
FROM OBSERVATIONS TO TEST DESIGN
We’re preparing three test pools: Basic for minimum expectations, Capability for differences between current systems, and Frontier for long-term challenges. As the field advances, tests can move between pools and the Basic standard will rise through new benchmark versions. These recordings inform test design; they are not scored benchmark runs.
Explore the three test tiersNEW EVIDENCE / OUTDOOR MOBILITY
Three excerpts add stair ascent, loss of balance, and hands-on assistance to the evidence archive. They inform candidate mobility tests; they are not scored benchmark runs. Stair difficulty and the test protocol must be defined before assigning a test pool.
The robot climbs the wide outdoor staircase and reaches the upper landing in this excerpt.
Source: DJI_20250911_153851_26_null_video.mp4, 0:05–0:13. Recorded September 11, 2025 on campus stairs. Unitree G1, teleoperated through velocity/heading commands with a separate locomotion stack. These are distinct attempts, not a controlled comparison with indoor HANDOFF runs.
A separate recording shows a loss of balance near the bottom of the stairs, followed by people intervening.
Source: DJI_20250911_154344_28_null_video.mp4, 0:21–0:35. Recorded September 11, 2025 on campus stairs. Unitree G1, teleoperated through velocity/heading commands with a separate locomotion stack. These are distinct attempts, not a controlled comparison with indoor HANDOFF runs.
A closer view at a different staircase shows a person making physical contact with the robot during the sequence.
Source: PXL_20250911_194213049.mp4, 5:05–5:20. Recorded September 11, 2025 on campus stairs. Unitree G1, teleoperated through velocity/heading commands with a separate locomotion stack. These are distinct attempts, not a controlled comparison with indoor HANDOFF runs.
Seven new recordings total approximately 24 minutes 29 seconds. Review used timeline sampling and closer inspection of selected sequences, not a complete trial audit. These recordings use different controllers and robot configurations from the May 2026 manipulation footage. No combined success rate is calculated.
Which stages of object handling are visibly achieved, where does balance break down, and what human assistance is visible?
01 / OBSERVATIONS FROM OUR ARCHIVE
SHELF SEQUENCE / APPROXIMATE SOURCE TIMES
Observed stages in one recording, not a success-rate measurement. Destination must be assessed against the task instruction; upright placement was not specified.
CASE STUDY 001 / RESEARCHER-CONFIRMED FINDINGS
1 kg box. Partial completion: the required turn and handover are not shown.
Correct object retrieved, transported and released at the specified destination. Task complete; clean success not verified.
Five steps, 80 cm ascent. Teleoperated completion; clean success not verified.
Six reviewed attempts: 2 task completions, 1 partial completion, 3 failures/interventions; 1 reset. This selected subset mixes tasks and configurations and excludes additional unaudited footage. It is not an overall robot success rate.
Full setup, task definitions and evidence ↗︎SIX DISTINCT ATTEMPTS / RETROSPECTIVE CASE STUDY
Confirmed outcomes, linked to each excerpt. Stage scores describe progress; they are not clean-success ratings. No overall reliability percentage is calculated across different tasks or controllers.
| Attempt / evidence | Control | Stage | Outcome | Timing | Assistance / reset |
|---|---|---|---|---|---|
| 01 · Box lift / hold ↗︎ | Autonomous · HANDOFF | 2 / 3 · lift | Partial; required handover not shown | ~6 s lift; ≥~15 s visible hold | 1 reset; clean success not achieved |
| 02 · Box balance loss ↗︎ | Autonomous · HANDOFF | 2 / 3 · lift initiated | Caught balance loss; intervention | Balance loss at 10 s | Recovery assistance; no ground-contact fall |
| 03 · Shelf retrieval ↗︎ | Autonomous · HANDOFF | 3 / 3 · placement | Task complete; clean status unverified | 60 s: first reach → completed release | Slack tether / no assistance not established |
| 04 · Stair ascent ↗︎ | Teleoperated · earlier stack | Not assigned² | Landing reached; clean status unverified | 4 s: first upward step → both feet stable | Slack tether / no assistance not established |
| 05 · Stair balance loss ↗︎ | Teleoperated · earlier stack | 1 / 3 · partial ascent | Caught balance loss; intervention | No complete-task time reported | Recovery assistance; no ground-contact fall |
| 06 · Stair assistance ↗︎ | Teleoperated · earlier stack | 1 / 3 · provisional | Hands-on assistance; completion not established | No complete-task time reported | Physical assistance recorded |
The hands-on stair score is provisional: partial ascent is documented but landing completion is not. ² The completed stair run remains unscored because support evidence is insufficient to distinguish assisted from clean completion. No clean-success designation is assigned to the completed shelf or stair run.
Archive summary: 2 task completions, 1 partial completion, 3 failures/interventions and 1 reset. The two balance-loss incidents were caught, not ground-contact falls. Additional unaudited footage exists; these six attempts are not ten consecutive attempts of one test.
CONFIRMED BY THE CONTRIBUTING RESEARCHER
These are retrospective observations of research demonstrations, not prospective trials of the proposed benchmark. The May manipulation and September stair recordings use different configurations.
RoboAnalyticag1_hand_moe_flat_unicmd_nobv_fullstable_seed_amp, exported from model_20000.pt. A VLM planner uses camera imagery and language instructions; task-space commands drive HANDOFF.Correction: the earlier blanket statement that no hand detached was incorrect. A separate stair attempt, DJI_20250911_155330_29_null_video.mp4 at approximately 00:37, visibly shows a left-hand component separating. This is distinct from the balance-loss attempt in the six-row results table; the exact detached assembly remains unconfirmed.
RESEARCHER-CONFIRMED RULES / PROTOCOL V0.1
0 No grasp
1 Grasp
2 Lift / transport
3 Correct placement / handover
Score the highest completed stage. Record assistance separately; a stage 3 result is not automatically clean.
0 No ascent
1 Partial ascent
2 Assisted completion
3 Clean completion
Teleoperated and autonomous runs remain separate, even when the same staircase is used.
At least 8 clean successes in 10 consecutive attempts. Assisted attempts stay in the denominator and count as failures. Recovery assistance does not disqualify the entire batch.
A slack fall-arrest tether is allowed. Lateral stabilization makes an attempt assisted and ineligible for clean success.
Report task completion, clean-success rate, time, intervention rate, reset count, ground-contact falls and caught balance losses separately. There is no invented penalty-weighted index. The present recordings do not establish a ten-attempt pass.
Shelf: first reach to completed release (60 s). Stairs: first upward step to both feet stable on the landing (4 s). Box lift: first task-directed reach to the box fully clearing its supporting surface. Box hold: clearance to first surface contact, handover, drop or clip end. The ~6 s lift and ≥~15 s visible hold are estimates, not verified measurements; clip-end censoring prevents treating the visible hold as a completed hold duration. Source excerpt: approximately 2:28–2:55 in PXL_20260515_001133772.mp4.
This procedure is a proposed implementation of the confirmed rules; it is not a claim that a new test series has been run.
Download ten-attempt recording sheet ↓02 / SUPPORTING PUBLICATION · AUTHOR-REPORTED RESULTS
HANDOFF combines planning and whole-body control on a Unitree G1. The researcher confirms that the shelf and box recordings match the published HANDOFF demonstrations and the checkpoint documented above.
Simplified explanatory diagram adapted from HANDOFF. Our evaluation observes the complete system’s behavior.
Retrieve the Pringles container and place it left of the Cheez-It box.
Lift the blue box, turn, and hand it over.
Source: HANDOFF project. Author-selected demonstrations; these are supporting research, not additional independently scored trials. Playback timing is as supplied by the authors.
PUBLISHED SIMULATION BENCHMARK / TABLE 2(b)
Feasible bilateral wrist-target trials (%) · higher is better
Feasible means both wrists within 15 cm, no fall, and pelvis drift below 25 cm. The study uses 2,000 discovery and 400 accuracy targets per controller (seed 42), with adapted interfaces for baselines. These are simulation controller results, not physical task success rates or robot-brand rankings.
Source: HANDOFF, Table 2(b) ↗︎The paper describes wrist-orientation and forward-camera limitations. Our next tests should vary grasp orientation and object visibility. Neither limitation establishes the cause of a particular failure in our recordings.
Research credit: Lizhi Yang and coauthors, HANDOFF, arXiv:2606.06493v1 (June 2026 preprint). Reuse authorized by the contributing cofounder. Published results have not been independently reproduced by RoboAnalytica.
Fix one task and configuration, define success before the test, retain all attempts, document assistance and support, and publish the outcomes with linked source evidence. Where simulation is used, match the controller and conditions and disclose material differences.
Help build the benchmark ↗︎