UPS Cut 20,000 Jobs Amid Labor Shortages. Its Fix Weighs 70kg and Sorts in 4 Seconds.

Figure AI's Helix model gained force feedback and temporal memory in June 2025, letting a humanoid recover from a conveyor jam or a coworker's bump instead of freezing mid-task. The upgrade coincided with renewed talks with UPS for warehouse trials, turned short customer deployments into continuous training-data capture, and arrived alongside a $1 billion Series C at a $39 billion valuation, Reuters reported — capital that tracks a logistics thesis.

The base Helix, introduced in February 2025, already outperformed typical robot controllers as a generalist vision‑language‑action model outputting continuous control of the full upper body (wrists, torso, head, and individual fingers) across a 35‑degree‑of‑freedom action space at 200 Hz. A slower onboard vision‑language model (System 2) ran at 7‑9 Hz for scene understanding, while a fast visuomotor policy (System 1) translated latent representations into actions. Training relied on roughly 500 hours of multi‑robot, multi‑operator teleoperated behavior. Impressive, but the system had no sense of touch and no recall of what it had just done.

The logistics-scale upgrade added vision memory, state history, and force sensing — described plainly as "touch" in a September 2025 recap of Figure's June 7 update. A new vision memory module gives Helix stateful perception; the policy now ingests a history of past states. That architectural tweak enables temporally extended behaviors and lets the robot shrug off interruptions that would have halted the earlier version. Force feedback entered through the state input. On the Figure 03 hardware, introduced late 2025, fingertip sensors detect forces as small as 3 grams, and palm cameras supply visual feedback when the main cameras are blocked (say, when the hand reaches deep into a bin). Together these additions let the model grip a fragile parcel without crushing it and read a barcode's orientation by feel rather than sight alone.

Brett Adcock said the milestone "is critical to unlocking the next stage of growth… scaling out our AI platform Helix." He also stressed the gain came from data, not new algorithms: "No new algorithms, no special‑case engineering, just new data." The company's own figures put the payoff in concrete terms during logistics scaling: handling dropped to about 4 seconds per package and barcode‑orientation success reached roughly 95%, signaling human‑level throughput. The robot that once needed a human to reset after a bump now feels the jostle and remembers the box it was holding.

UPS Pilot: From Demo to Dock

Talks between Figure AI and UPS began in 2023 and resumed in early 2025 after the Figure 02 demonstrated significant performance gains, per Bloomberg. The renewed dialogue follows Figure's first confirmed commercial deployment: a gradual, long-term collaboration with BMW at the automaker's Spartanburg, South Carolina plant, where robots have assisted with material handling since 2024. On that line, Figure 02 achieved a 400 percent speed increase over its initial cycle times — a data point that carried weight when conversations with UPS restarted.

The robot under discussion stands five feet six inches tall, weighs roughly 70 kilograms, and lifts up to 20 kilograms (55 pounds). Its hands offer 16 degrees of freedom. Designed to operate in environments built for people, it performs repetitive lifting, sorting, and transport tasks while responding to spoken commands. A February video posted by Figure AI showed the 02 beside a conveyor belt, picking and sorting small parcels, a direct preview of the logistics triaging the company now describes as "a new real-world application for Figure robots: logistics package manipulation and triaging," adding that the task "demands human-level speed, precision, and adaptability, pushing the boundaries of pixels-to-actions learned manipulation."

UPS is no stranger to automation. The carrier has deployed more than 700 robots in its Louisville, Kentucky hub and other locations, and runs fixed robotic arms alongside AI-driven software in its high-tech Velocity facilities. It also partnered with Dexterity Inc., a startup building industrial robots capable of "human-like" finesse. But humanoid robots represent a different category: mobile, bipedal systems that can navigate the same aisles, stairs, and work cells as human workers without requiring conveyor redesign or cage infrastructure.

The initiative remains exploratory. UPS has not disclosed how many robots it might test or when trials could begin, though internal planning is under way. If finalized, the deal would make UPS one of the first global logistics firms to pilot humanoid robots at scale. The exact functions Figure's humanoids would handle for UPS remain undefined, but the carrier's stated evaluation criteria focus on integration into existing workflows and whether the technology can improve efficiency and address labor shortages in key facilities.

Those shortages are measurable. Roughly three in four logistics businesses report staffing gaps. UPS itself is cutting 20,000 jobs and closing 73 facilities by June 2025 to save $3.5 billion, pressured by falling package volumes, rising operational costs, and a tough global economy. Manufacturing and logistics face high turnover, safety risks, and rising wage pressure. UPS has been spending around $1 billion annually on automation and AI to cut costs and boost efficiency — a budget that now includes humanoid evaluation.

If the pilot proceeds, UPS would join a growing list of major corporations experimenting with AI-powered robotics to future-proof operations. Humanoid robots that safely lift packages, navigate facilities, and interact with human supervisors could significantly reduce reliance on temporary or seasonal workers. They could offer a more flexible solution compared to traditional automation tools, performing a variety of tasks without requiring significant reprogramming or additional infrastructure.

Skepticism exists. Fortune reported in April 2025 that some observers questioned whether Figure may have exaggerated the extent of its work with BMW AG. But the trajectory is clear: experts see humanoid robots going mainstream in logistics within five to ten years. Figure plans to have robots working side-by-side with humans, or independently, in factories, warehouses, and eventually homes. With tech giants and billionaires backing it, Figure AI may have the clearest path to commercial deployment. If UPS finalizes its pilot deal, it could set the standard for how robots and humans share space on factory floors and in shipping hubs. Other logistics and manufacturing giants are expected to follow closely once one major player validates humanoid labor at scale.

The Data Engine: Why Six-Month Roles Matter

Figure AI's active humanoid fleet passed 700 robots in early July 2026, outnumbering the company's roughly 650 employees. That hardware count means little without the training streams that teach Helix to manipulate packages. Data is the bottleneck for embodied AI, not compute, as EV Sint wrote in April 2026. The logistics push, including BMW's 11-month Figure 02 run and the ongoing UPS warehouse talks, turns short customer deployments into collection windows. A six-month slot on a plant floor is not a finished product launch. It is a data haul.

The BMW Group Plant Spartanburg deployment shows the pattern. Figure 02 moved from initial setup to plant testing within six months in 2025 and reached full deployment on an active line within 10 months, running every working day. Across that stretch the robot loaded more than 90,000 parts, logged more than 1,250 hours of runtime, and contributed to over 30,000 BMW X3 vehicles. Those hours are the raw material. Helix does not learn from a static dataset pulled from the web. It needs synchronized action-state streams captured while a robot actually grips and rotates a package for scanning.

Figure AI's own logs show how thin the starting data was. In February 2025 the company reported that just 8 hours of well-curated demonstration data yielded a dexterous policy for package rotation. Three months later, scaling training demonstrations from 10 to 60 hours cut average processing time per package from ~6.84 seconds to 4.31 seconds and lifted barcode success from 88.2% to 94.4%. The jump came from both data scaling and model architecture, but the company is blunt: for a single use case, data quality and consistency matter more than quantity.

The numbers above are from Figure AI's 2025 logistics reports. They show why a six-month customer role is a strategic asset. A robot on a line for a quarter or two generates the demonstration hours that turn a hesitant grip into a 4-second scan. Force sensing integrated into the state input adds another stream that must be recorded alongside vision memory, but the capture problem stays the same: embodied policies need paired action and state at scale.

That logic drives Figure's internal staffing. Zero G Talent's board lists 77 open roles at Figure AI, with 11 added in the past week. The newest posts include Helix AI Engineer, Android and Helix AI Engineer, Backend Infrastructure, both paying $150,000–$400,000 a year, plus Software Engineer, Privacy & Data Governance at $150,000–$350,000. The median band across the board sits at $200,000. These are not assembly-line jobs. They build the pipeline that ingests robot telemetry and labels it for retraining.

Project Go-Big, Figure's large-scale humanoid pretraining data-collection initiative, depends on keeping hundreds of robots productive at customer sites rather than in a lab. The company reported an end-of-line first-pass yield above 80% for its Bot Q production, meaning four of five robots passed final test without rework. But the Figure 03 sequencing project remains early stage and should not be called mature production. The data capture engine must run continuously across both research units and commercial trials.

Apptronic's Apollo 2 offers a parallel path: a modular platform deliberately built to collect real-world data for its next commercial robot. Apptronic uses teleoperation and customer testing with Google DeepMind's Gemini robotics research to feed Apollo 3. Figure's bet is different. It places bipedal robots directly into logistics flows where full-body coordination matters, and treats each shift as a labeling session.

The six-month role is therefore a temporary deployment with permanent residue. When BMW's line went live with Figure 02, the robot ran 10-hour shifts Monday through Friday and hit more than 99% successful placement per shift. Those shifts ended, but the policy improvements stay. UPS trials will open another such window. The warehouse floor is the new annotation queue, and the clock starts the moment a robot picks up its first box.

Capital Bets on Robotic Fulfillment

Figure AI closed a $675 million Series B in February 2024 and later added a $1 billion Series C, lifting total capital raised to $1.9 billion at a $39 billion valuation. The backing did not come from sci-fi hype about home helpers. It tracked a logistics thesis: put humanoid robots into warehouses and supply chains.

Figure built that thesis into its funding narrative from the start. "Our Humanoid is designed for initial deployment into the workforce to address labor shortages, jobs that are undesirable or unsafe, and to support supply chain on a global scale," the company said in its February 2024 raise announcement. Investors bought the pitch. Microsoft, OpenAI Startup Fund, NVIDIA, Jeff Bezos via Bezos Expeditions, Intel Capital, Parkway Venture Capital, Align Ventures, and ARK Invest put money into the Series B at a $2.6 billion mark.

NVIDIA appears in both rounds. That dual check matters more than a single logo on a press release. The chip designer does not write checks to humanoid startups that plan to train models on someone else's silicon. Figure stated the Series B would accelerate commercial deployment, and Forbes reported the $1 billion raise aimed to "build AI infrastructure, capture training data and make robots." AI infrastructure in 2024 to 2025 means GPU clusters, almost certainly NVIDIA's. The logistics vision needs force feedback and temporal memory models, and those models burn compute. Figure also confirmed it would use Microsoft Azure for AI infrastructure, training, and storage; Azure's AI supercomputing tier runs on NVIDIA hardware. The chip buy is thus indirect but real: NVIDIA's capital and its silicon are both inside Figure's stack.

The capital inflow lined up with UPS's own crisis. Bloomberg told Yahoo Finance in May 2025 that UPS and Figure first talked in 2023, resumed talks early 2025 as Figure's model hit new milestones, and that UPS planned to cut 20,000 jobs and close 73 facilities to save $3.5 billion. Falling package volumes from Amazon and tariff pressure pushed the shipper toward automation. A humanoid that lifts 55 pounds and triages parcels suddenly looked like a balance-sheet line item, not a lab curiosity. Reuters via Yahoo noted UPS faces falling volumes, rising costs, and a tough global economy. Figure's raise timing rode that pain.

Competitors in the humanoid race have not matched the war chest. Tesla's Optimus and Agility Robotics' Digit are vying for the same warehouses, but Squaredtech reported in July 2025 that neither had raised as much capital this quickly. Figure's $1.9 billion total gives it runway to place robots inside real flows rather than film demo clips. Valuation jumped from $2.6B to $39B in roughly a year, a signal that investors price logistics deployment higher than home novelty.

The money is moving into heads, not just hardware. Figure AI posted 11 roles in the past 7 days, with a median salary of $200,000 across 77 open listings. The newest include Helix AI Engineer and Backend Infrastructure roles in San Jose paying up to $400,000. That hiring surge follows the Series C and signals that the logistics build-out is staffed. The company's own site repeats the goal of commercial operations "as soon as possible."

The bet is not abstract. NVIDIA's repeated checks and Figure's Azure plan put real silicon behind the Helix model. If UPS finalizes a pilot, the next investor signal will be revenue, not raise.

Scope Limits: What This Is Not

Figure's Helix upgrade and the UPS trial talks that followed are a logistics story — not a defense story, not a home-care story, and not a demo-story. The distinction matters because capital, talent, and regulatory scrutiny flow differently across each front, and conflating them obscures what Figure is actually building.

Start with defense. The Department of Defense has been fielding AI-enabled systems for over 60 years; its 2023 Data, Analytics, and AI Adoption Strategy supersedes a 2018 AI Strategy and a 2020 Data Strategy to accelerate adoption "at the speed of relevance and at the scale of our global mission." The European Defence Fund's 2025-2027 perspective prioritizes trustworthy AI for "adversarial high-risk scenarios, demanding extreme reliability, resilience and security" — stringent testing and robust cybersecurity are mandatory, not optional. Figure's CEO has said plainly: "We don't spend too much time looking at competitors" and "We don't go to a lot of events. I think it's a giant waste of time." The company's prime directive, stated on the record, is "are you here to sell work? We're selling work." That work is package triaging at UPS and sheet-metal handling in BMW's Body Shop — not missile defense, not battlefield autonomy.

Home care is a separate front. Figure's own site describes Helix as AI that "enables it to navigate unpredictable, ever-changing home environments," yet the company's deployed fleet runs in a Body Shop moving sheet metal back and forth and in a second logistics customer's facility. The CEO has emphasized consumer electronics manufacturing, not car manufacturing, and a path to 100,000 units in four years built on "the experience curve of manufacturing" and fleet learning — not on bedside assistance.

UX-only humanoid narratives — robots that talk, gesture, and stage impressive demos but ship zero productive hours — are the third exclusion. Boston Dynamics brought a robot to the same Bloomberg Tech summit where Figure appeared. Agility Robotics' Digit has the most warehouse deployment experience via its Amazon partnership. But Figure's Helix architecture adds temporal memory, including a vision memory module giving stateful perception and history-dependent policies, and force feedback integrated into the state input, specifically for "pixels-to-actions learned manipulation" at production speed. The CEO's framing: "There's no remote control or teleoperation of your Figure 02 robot. We do none of this in real autonomous activities. We're trying to deliver a full stack solution vertically integrated that can go in and autonomously do work with just speech." That is a claim about operational autonomy in a controlled logistics envelope, not a claim about general-purpose embodied AGI.

The $675M–$1B raise and Nvidia chip buys are priced to that logistics scope. The 11 roles Figure posted in the past week, including Helix AI Engineers, Security Engineers, and Privacy & Data Governance, map to hardening a production stack for warehouse deployment, not to clearing defense export controls, not to HIPAA compliance, not to demo-day choreography. The fence is deliberate: Figure is building a robot that earns its keep sorting packages. Everything else is a different company's roadmap.

The warehouse floor is the annotation queue. The robot picks up a box, feels its weight, reads the barcode by touch when vision fails, and remembers the grip when a conveyor shudders. That loop — sense, act, remember — is the product. The rest is noise.

Working in robotics? Zero G Talent tracks the openings: see every open Figure AI role, browse robotics jobs, openings at Temporal Technologies and Leap, and the people building the field.