Figure Helix-02 : humanoid robots complete full 8-hour shifts in a factory — without human intervention

Figure Helix-02: humanoid robots complete full 8-hour shifts in a factory — without human intervention

🔎 On May 13, 2026, humanoid robotics stopped being a laboratory spectacle

Brett Adcock, CEO of Figure AI, posted a simple tweet on May 13, 2026. Below it: eight hours of livestream showing a team of humanoid robots sorting packages on conveyors, without any human intervening. No resets, no hidden teleoperator, no technical stoppage.

The video showed Helix-02 robots operating at performance levels equivalent to a human operator, according to Figure AI's statements. This is the first time a manufacturer has claimed continuous industrial autonomy over a standard work duration.

What changes the game is the total absence of human supervision during those eight hours. Previous industry demonstrations — from Tesla, Boston Dynamics, or Agility — remained limited to carefully edited sequences of a few minutes. Here, Figure AI played the card of raw transparency.

The global context reinforces the impact: the logistics labor shortage is worsening in developed economies, and Chinese competitors like Unitree are already deploying their own humanoids in public environments. The question is no longer "if" but "when" large-scale production will begin.

The essentials

• On May 13, 2026, Figure AI demonstrated a complete 8-hour industrial shift performed by Helix-02 humanoid robots, without any human intervention (source: TechTimes).
• Helix-02 is Figure's second-generation AI system, distinct from Helix v1 which targeted domestic tasks. It relies on a "hive-mind" with shared neural weights.
• Helix-02's neural networks replaced 109,504 lines of classic code, enabling self-diagnosis and real-time maintenance coordination.
• The Figure 02 robot weighs 70 kg, measures approximately 170 cm, and features a 2.25 KWh battery — double the previous generation (source: Built In).
• Figure 03, the next model, targets the domestic market at ~$600/month, while Helix-02 remains industrial-focused — two parallel strategies.

Recommended tools

What Helix-02 actually does — and what it doesn't do

Helix-02 wasn't just sorting random things for eight hours. The demonstration focused on conveyor parcel sorting: picking up, identifying, and placing items into the correct bin. This is a highly standardized logistics use case, not unstructured manipulation.

The technical key is whole-body robot control. Unlike approaches that separate locomotion from manipulation, Helix-02 drives the entire body as a unified system. The robot adjusts, leans, and recalibrates in real time — much like a tired human compensating for fatigue.

Multi-minute control is the other substantial breakthrough. Previous systems would lose track after a few dozen seconds, requiring recalibration. Helix-02 maintains task consistency over cycles lasting several minutes, strung together uninterrupted for eight hours.

What it doesn't do: it cannot handle the unexpected outside of its training domain. An unusually shaped package, a conveyor belt that stops, an obstacle on the floor — these scenarios fall outside the scope validated by Figure AI during this demonstration.

It also doesn't replace human preventive maintenance. The robot self-diagnoses and reports its needs, but a technician is still required for physical interventions on the hardware. This is a point often glossed over in the company's communications.

The distinction with Helix v1, Figure AI's initial system that targeted household tasks, is fundamental. Helix-02 is a complete overhaul, optimized for industrial repetition and endurance, not for domestic versatility.

109,504 lines of code replaced by neurons — what this means in practice

The figure is striking: Helix-02's neural networks replaced 109,504 lines of traditional code, according to Gadget Review's analysis. This reveals Figure AI's architectural philosophy.

Traditional robotic code operates on explicit rules: "if the package is on the left, move the arm X degrees." This is fragile, hard to maintain, and incapable of adapting to real-world variations. Every new type of package, every change to the conveyor belt requires an update.

By switching to a fully neural system, Figure AI trades explicit control for generalization. The robot learns to seamlessly associate perception and action, without every micro-decision being hard-coded.

The most important practical consequence is self-diagnostics. A rule-based system doesn't know it's malfunctioning until the malfunction triggers an explicit error. A neural network, on the other hand, can detect anomalies in its own internal signals and flag a need for maintenance before a breakdown occurs.

Coordination between robots is the other benefit. The "hive-mind" described by New Atlas means that all robots share the same set of neural weights. A robot that learns to better grasp a certain type of package transmits this improvement to the others — in theory. Figure AI has not detailed the real-time synchronization mechanism.

The risk of this approach: opacity. When 109,504 lines of code produce a bug, you trace the faulty line. When a neural network makes a bad decision, you analyze gradients and activations. Debugging is radically more complex, and this is a major issue for industrial adoption where legal liability is involved.

Figure 02 : the hardware behind the performance

Software does nothing without a body capable of supporting it. The Figure 02, which runs Helix-02, deserves a closer look because its specifications explain a large part of the demonstration.

2.25 KWh battery, double the previous generation, according to Built In. It is this capacity that theoretically makes an 8-hour shift possible. For comparison, most competing humanoids operated in the 1 to 1.5 KWh range in 2024-2025, limiting their autonomy to 2-4 hours.

70 kg for ~170 cm in height, reported by Interesting Engineering. This falls within the range of an adult human, which is not insignificant: it means that workstations designed for humans do not need to be redesigned.

6 computer vision cameras, offering near-complete stereoscopic coverage of the environment. Sorting packages requires precise 3D perception to assess the size, position, and orientation of each object.

4th-generation hands, a critical point that is often underestimated. Handling packages requires adaptable grasping — a smooth cardboard box is not grabbed the same way as a plastic bag. The quality of the hands directly determines the success rate of grasps.

Speech-to-speech reasoning system, interesting but secondary for the industrial use case. It could, however, be used for interaction with human supervisors on the line.

The Figure 02 is commercial-ready according to New Atlas, batteries included. This detail ("batteries included") is not insignificant: in the industry, the battery is often sold separately or offered as an option, inflating the real cost of deployment.

The comparison with the other contenders in the humanoid race shows that Figure AI has opted for complete vertical integration: hardware, software, AI. This is an advantage in terms of consistency, but a risk in terms of development costs.

8 hours without intervention — decoding the claim

This is the core of the announcement, and it needs to be dissected carefully. "A full 8-hour shift without human intervention" does not mean "8 hours without any issues."

According to Humanoid Guide, Figure AI claims that robots can carry out 8-hour autonomous industrial shifts. The May 14, 2026 livestream, described by FaithTechate, showed a "fully autonomous humanoid operation."

But "without intervention" concretely means: no human touches the robot, recalibrates it, or restarts it. It does not mean zero picking errors, zero missed packages, or zero micro-stops. A robot that misses a package and recovers on its own is still "without intervention."

The fact that Brett Adcock announced the demonstration via a simple tweet, without a structured press release, is strategic. It allows control of the narrative without exposing oneself to difficult technical questions. The video speaks for itself — or rather, that's what Figure AI wants people to believe.

Polymarket, the predictive betting platform, quickly created markets on the failure-free operating time of the Helix-02. This is an indicator of market confidence: bettors are not betting on "does it work," but on "exactly how long." The skepticism is about the duration, not the principle.

The honest comparison: a human operator at a package sorting station processes between 200 and 600 packages per hour depending on the complexity. Figure AI did not publish an hourly throughput figure for this demonstration. "Performance equivalent to humans" is a qualitative claim, not a quantitative one. As long as we don't have the number of packages processed per hour along with the associated error rate, the comparison remains incomplete.

From pilot to production: where does Figure AI actually stand?

Figure AI already deployed robots at BMW last year, reminds Interesting Engineering. This is an important fact: the company is not at its first deployment in an industrial environment.

But there is a major difference between a pilot at BMW — where the robot performs a few specific tasks in a controlled environment, with engineers on site — and an autonomous 8-hour shift in logistics. The first is a proof of concept, the second is a claim of operational maturity.

Figure AI's two-pronged strategy is ambitious. On one hand, Helix-02 for industry: long shifts, repetitive tasks, calculable ROI. On the other hand, Figure 03 for the domestic market at ~$600/month: a home robot, a mass market, a radically different value proposition.

Targeting both markets simultaneously is a risky bet. Tesla made the same choice with the Optimus (industrial then domestic), and had to scale down its ambitions on the timeline. The constraints are different: the industrial sector tolerates a 70 kg robot, the domestic market wants something safe around children. Industry pays for ROI, the domestic market pays for convenience.

The BMW deployment remains Figure AI's only publicly documented customer use case to date. A pilot does not make a business model. To scale up, you need volume contracts, a maintenance network, contractual SLAs — everything that makes the difference between an R&D lab and an industrial company.

Chinese competition, with players like Unitree already deploying its G1 in airports in Japan, is advancing on the ground of real-world deployment. The Chinese ecosystem benefits from a more mature robotics supply chain and potentially lower production costs.

What general AI models don't explain

It is tempting to link Helix-02's performance to the progress of general LLMs. After all, the best agentic models like GPT-5.5 (agentic score: 98.2) or Gemini 3 Pro Deep Think (95.4) show impressive reasoning and planning capabilities.

But the reality is more nuanced. Helix-02 probably doesn't use a generalist LLM in real time for every sorting action. The latency of a model like Claude Opus 4.7 or GPT-5.5, even when optimized, would be incompatible with the real-time constraints of motor control at 100+ Hz.

What is more likely happening: LLMs are used offline for the training and task planning phase, and then the control policies are distilled into lighter and faster networks that run locally on the robot. This is the standard approach in the industry — but it is rarely explained clearly in marketing communications.

The speech-to-speech reasoning of the Figure 02, on the other hand, could very well call upon a model of the caliber of GPT-5.4 or Claude Sonnet 4.6 for understanding and generating responses. But that is a peripheral module, not the core of the control system.

The common mistake in tech journalism is confusing "the robot uses AI" with "the robot runs GPT-5.5 every millisecond". Helix-02 is a specialized AI system for robotic control, not a chatbot on legs.

Implications for factory and warehouse workers

TechTimes analyzes the direct implications for warehouse and factory workers. We must tackle this issue head-on, without dramatizing or minimizing it.

Sorting packages on a conveyor belt is a high-turnover job, physically demanding, with high injury rates (lower back pain, musculoskeletal disorders). It is also a job where the labor shortage is most acute in developed countries. Logistics companies are struggling to recruit, even when raising wages.

In this context, a robot capable of holding a sorting position for 8 hours without a break, without sick leave, and without injury, represents a powerful economic argument. The calculation is simple: robot cost + maintenance vs. wage cost + employer contributions + turnover + sick leave.

But the replacement will not be binary. Where a warehouse employs 50 sorters, it will not buy 50 robots overnight. The realistic scenario is a gradual mix: 10 robots for the most repetitive positions, 40 humans for exception handling, supervision, and maintenance.

The most threatened jobs in the long term are those with very low cognitive added value: sorting, simple transport, basic palletizing. The protected jobs are those that require judgment, negotiation, crisis management — everything a Helix-02 does not do.

The social question is not "will we be replaced" but "at what speed and with what transitions." Figure AI does not publish any social impact study, which is revealing. The responsibility for the transition lies with the states and the user companies, not the manufacturer — but the latter's silence is notable.

❌ Common mistakes

Mistake 1: Confusing 8-hour autonomy with 8-hour reliability

"Unassisted" does not mean "error-free". A robot that fails on 5% of packages but recovers on its own is technically "unassisted". Reliability is measured in MTBF (Mean Time Between Failures) and success rate per task, not in time spent on the line. Figure AI has not published any of these metrics for this demonstration.

Mistake 2: Extrapolating beyond the demonstrated use case

A robot that sorts packages on a conveyor for 8 hours does not know how to assemble an engine, clean a workbench, or handle a complex unexpected event. Each task requires specific training. Generalization exists in models, but its limits remain unclear. Do not confuse an "impressive demo" with a "universal robot".

Mistake 3: Ignoring the total cost of ownership

The purchase or rental price of a humanoid is just the tip of the iceberg. Preventive maintenance, software updates, charging infrastructure, on-site integration engineering, team training — the TCO (Total Cost of Ownership) is typically 2 to 4 times the purchase price over 5 years. No TCO figures have been communicated by Figure AI.

Mistake 4: Comparing with resting humans

An 8-hour shift for a robot is not comparable to an 8-hour shift for a human. Humans take micro-breaks, naturally adapt to fatigue, and compensate with experience. The robot, on the other hand, maintains a constant level — which is an advantage for consistency, but a disadvantage if it enters a degraded state without detecting it.

❓ Frequently Asked Questions

Is Helix-02 the same system as Figure AI's initial Helix?

No. Helix v1 targeted domestic tasks with a generalist approach. Helix-02 is an industry-oriented redesign, optimized for endurance, repetition, and whole-body control. Both share a neural philosophy but differ in their architecture and targeting.

What is the relationship between Figure 02 and Figure 03?

Figure 02 is the current, commercial-ready robot used for industrial deployments with Helix-02. Figure 03 is the model announced for the domestic market at ~$600/month. These are two distinct products for two distinct markets, even though they share technological foundations.

Can this really be compared to competitor deployments like Unitree's?

Partially. Unitree deploys its G1 in public environments like Haneda Airport, which proves real deployment capability. But the use cases are different: reception and navigation for Unitree vs. intensive manipulation for Figure. A direct comparison is difficult, but both indicate that humanoid robotics is leaving the labs.

Is an LLM like GPT-5.5 or Claude Opus 4.7 used in Helix-02?

Probably not in real time for motor control. The latencies of general LLMs are incompatible with high-frequency control. However, they could be involved in offline planning phases or in the speech-to-speech module. The core of the control relies on lighter, specialized networks.

Does this mean the end of warehouse jobs?

Not in the short term. The realistic scenario is a progressive human-robot mix, with humanoids taking over the most repetitive and physically demanding tasks. Low cognitive-value jobs are the most exposed in the medium term, but the transition will depend on costs, regulations, and companies' ability to integrate these systems.

✅ Conclusion

Figure AI has crossed a symbolic milestone with Helix-02: moving from a few-minute demonstration to claiming a complete industrial shift. The Figure 02 hardware and Helix-02's neural architecture make the scenario credible, but reliability and cost metrics remain opaque. The race for humanoid robots is no longer a question of technical feasibility — it has become a question of economic viability and industrial integration. The 8 hours without intervention are a strong signal, not formal proof of maturity. The market will judge based on upcoming deployment contracts, not on livestreams.