Remote Experts, Helix 02 and the Missing Metric Behind Humanoid Autonomy Claims

Humanoid autonomy is not one binary state. A robot can be teleoperated continuously, remotely assisted only after failure, run a scripted sequence, execute a learned policy after a human starts the task, or sustain long-horizon behavior with autonomous recovery. Those modes have different operational value and different data requirements.

For this article, autonomous evidence means the robot performs the physical task after start without a human steering its motion during the cited sequence. Remote assistance is recorded separately. Data collection by teleoperation is also separate from the control mode used at deployment time.

1X's current NEO pages state that the robot arrives with basic autonomy for early owners and uses Redwood AI for learning and repeating tasks. For chores it does not know, a 1X Expert can remotely supervise its actions at scheduled times. 1X also lists remote control through the mobile app and VR device as a separate feature.

That disclosure is valuable because it makes the service boundary visible. A completed home chore does not by itself prove that every motion was generated without remote assistance. To measure progress, 1X would need to publish how often Expert Mode is invoked, for which chore classes, how long interventions last, and how often the robot later completes the same chore without assistance.

Figure describes Helix 02 as a hierarchy from semantic goals to whole-body control. System 2 handles scene understanding and language-level sequencing. System 1 maps onboard sensing to full-body joint targets at 200 Hz. System 0 tracks whole-body motion at 1 kHz and was trained using more than 1,000 hours of retargeted human motion plus simulation-based reinforcement learning, according to Figure.

The strongest cited Helix 02 demonstration is the four-minute dishwasher sequence published January 27, 2026. Figure says the robot completed the end-to-end task with no reset and no human intervention and describes 61 loco-manipulation actions. That is meaningful evidence for one continuous task. It still does not reveal a success distribution across many kitchens, object sets, robot units and repeated trials.

On June 30, 2026, Figure showed Figure 03 at BMW Group Plant Spartanburg performing a logistics sequencing workflow under Helix 02. The robot handled thin-walled parts, repositioned its body while manipulating, and pulled a cart. The company describes this as a Figure 03 manufacturing demonstration. The public post does not publish a Figure 03 intervention rate, cycle-time distribution or long-duration runtime for that workflow.

Figure's earlier Figure 02 deployment is useful because it shows what mature reporting can look like. Figure reported 1,250+ runtime hours and 90,000+ parts loaded during an 11-month BMW program, and it named three KPIs cycle time, placement accuracy and human interventions. Those numbers cannot be transferred to Figure 03, but they show the operational metrics that should accompany future Figure 03 claims.

A fair autonomy comparison needs a denominator. Report successful tasks per 100 attempts, interventions per 100 tasks, interventions per operating hour, median and 95th-percentile task time, reset count, autonomous recovery rate, and performance by task class. Keep remote assistance time separate from autonomous execution time.

Long unedited sequences are more informative than highlight reels, but even an uncut video proves only that the shown sequence occurred. Repeated trials across different homes, object arrangements, lighting, floor conditions, network states and robot units are needed to estimate reliability.

Home robots face clutter, reflective or transparent objects, deformable items, pets, narrow passages, stairs, uncertain object placement, wireless outages and privacy constraints. A hybrid service model also adds response-time, network and remote-support dependencies. The useful question is what the robot does when confidence is low and whether the fallback is safe and visible to the user.

Factory humanoids face different constraints takt time, fixture tolerance, part variation, battery scheduling, repetitive loading, thermal limits, maintenance, calibration and restart after safety stops. A robot can be autonomous at the policy level yet still require frequent technician intervention for hardware, calibration or cell faults.

For every demo, write down the exact robot version, software version, task, environment, duration, number of attempts, success definition, human control mode, reset count and recovery behavior. If any field is missing, leave it missing rather than estimating it from the video.

For a product purchase, add service terms, remote-support conditions, privacy controls, software rights, connectivity requirements, maintenance, spare parts and what happens when the robot cannot complete a task. Autonomy is only useful when the fallback and support model are acceptable in the intended environment.

The most useful next disclosure from 1X would be chore-level autonomous completion and Expert Mode invocation rates over weeks of real home use. The most useful next disclosure from Figure would be Figure 03 intervention-normalized reliability and cycle-time distributions from repeated BMW or home operation.

A shared third-party protocol would be stronger still same task families, clear starting conditions, unedited trials, published failures, fixed intervention definitions and enough repetitions to show variance. Until then, the defensible comparison remains evidence-by-task rather than one autonomy leaderboard.