2026-05-13
Figure AI live-streams an 8-hour unbroken Helix-02 shift on a packaging line
— views
Figure AI broadcast an unedited 8-hour livestream of humanoid robots running Helix-02 on package-sorting conveyors. Parallel deployment at BMW Spartanburg has now contributed to 30,000 cars.
On May 13, 2026, Figure AI CEO Brett Adcock posted to X: “Watch a team of humanoid robots running a full 8-hr shift at human performance levels. This is fully autonomous running Helix-02.” The accompanying broadcast — running unedited for the full eight hours — showed Figure 02 humanoids on a package-sorting conveyor cell, reading barcoded packages and routing them at a pace Figure says matches a human sortation worker. It is the first long-form public livestream of a humanoid foundation model running production-grade work without on-screen human intervention.
What the livestream actually shows
The livestream cell is not the BMW deployment (that is a separate, ongoing production-line program). The 8-hour stream is a Figure-controlled sortation testbed:
- Pick-and-place from an inbound chute, scan barcode, route to one of N downstream bins
- Repeats over an 8-hour shift with no operator entries
- Multiple Figure 02 robots in the same cell, all running shared Helix-02 weights
- Per-robot uptime and pick-rate displayed on screen
The value of the eight-hour duration is boring data: stuck grippers, slip events, recovery behavior, hand wear, sensor noise drift. A 30-second demo can hide all of these; eight unedited hours cannot.
Helix-02 in one paragraph
Helix-02 is Figure’s vision-language-action (VLA) foundation model, a unified neural network that handles walking, manipulation, balance, and whole-body coordination from a single policy. Unlike dual-system designs (a VLM that talks to a separate low-level controller), Helix-02 combines vision, touch, proprioception, and whole-body control into one learning system designed for long-horizon tasks. The model is shared across all robots in the cell — there is no per-robot fine-tune, and a single weight update propagates to every unit.
The parallel BMW Spartanburg deployment
While the livestream is Figure’s own facility, the company has been running Figure 02 humanoids on a production line at BMW Group Plant Spartanburg in South Carolina. Per Figure’s own published numbers, the robots have so far contributed to the production of 30,000 BMW vehicles and moved more than 90,000 parts. Earlier disclosures referenced 10-hour shifts at BMW facilities. This is the most quantitatively-disclosed humanoid factory deployment to date.
Why the 8-hour livestream matters
Figure has been criticized — by Brett Adcock himself in past public remarks — for cherry-picked demo videos. The eight-hour unedited broadcast is the response. A live feed makes it expensive to hide failure modes: stuck grippers, dropped packages, unrecoverable poses, the model just sitting motionless because it lost the task. The per-robot dashboard is the meaningful artifact, more than any single successful pick.
The cynical read is that this is marketing positioning ahead of further fundraising and the Figure 03 ramp. The generous read is that it is also the first time the public can audit a humanoid deployment’s actual numbers in real time rather than waiting for a quarterly customer testimonial.
What practitioners should watch
The stream’s value depends on what you’re trying to learn:
- If you build VLA models: watch the failure recovery behavior. Helix-02’s design hypothesis is that a single end-to-end policy can learn graceful failure recovery without explicit fallback rules. If you see robots reattempting failed grasps in novel ways (not just retry-same-pose loops), that’s evidence for the hypothesis.
- If you build humanoid hardware: watch the hand and wrist wear. The Figure 02 hand is a multi-DOF design with under-actuated fingers. Eight hours in is when you’ll see if the hand drifts, cables stretch, or pads wear under repeated contact.
- If you build manufacturing software: the BMW Spartanburg numbers (30K vehicles, 90K parts) matter more than the sortation livestream. Compare the Figure-Helix-02 sortation throughput against incumbent automation (Symbotic, AutoStore, Berkshire Grey alternatives) for the same SKU mix.
What this does NOT prove
An 8-hour run is not 6 months of mean-time-between-failure data. Even a flawless stream tells you the system can run for 8 hours in a known configuration — it tells you nothing about a model drifting under monthly weight updates, sensor degradation, lighting changes through the year, or what happens when SKU mix shifts. Real industrial deployment proof points need quarterly statistics and adversarial testing reports. Treat the livestream as a marketing artifact backed by real data, not a maturity claim.
Where this fits in the May 2026 humanoid landscape
For context within the field:
- Tesla Optimus — reportedly running on a Tesla line, but no livestream and no published numbers comparable to Figure’s BMW disclosure.
- Apptronik Apollo — on a Mercedes line at Berlin (logistics tasks); the company prefers periodic press tours over live data.
- 1X NEO — pivoted to home pre-orders earlier in 2026; no published industrial deployment.
- Agility Robotics Digit — has the longest commercial runway (warehouse logistics since 2024) but the use case is wheeled-base logistics, not bimanual manipulation.
Figure’s bet is that end-to-end VLA learned from teleoperation generalizes faster than rule-coded systems. The BMW Spartanburg numbers and the 8-hour livestream are the two strongest public data points it has to support that bet today.