Robotics Weekly Review 2026-06-07

Week In Review

This week underlined a shift the field has been hinting at for a year: humanoid robotics moved decisively from demonstration to platform. NVIDIA and Unitree jointly unveiled the Isaac GR00T Reference Humanoid Robot, an open, off-the-shelf research stack built on the Unitree H2 Plus chassis; a day later Boston Dynamics’ Atlas appeared in Hyundai’s “School of Football” film, executing soccer-skill demonstrations learned from human motion data. Both stories are about scaling — for NVIDIA, scaling the developer community by collapsing the cost of starting humanoid research, and for Boston Dynamics and Hyundai, demonstrating that reinforcement-learned full-body skills now transfer cleanly from human capture to a 1.5-meter electric biped.

Industry deployments matured in parallel. Amazon launched its next-generation Proteus at a Delivering the Future event in London, with conversational language as the new control surface and operations expanding beyond dock areas to entire fulfillment centers. BYD confirmed it is developing humanoids and plans to deploy 20,000 units inside its own factories this year, joining Figure, Tesla, and Apptronik in the small group of companies treating humanoids as both product and production tool. Figure’s BotQ facility reached a milestone of one Figure 03 robot rolling off the line every hour, a 24-fold throughput increase in 120 days that reframes “supply-constrained” from a technical limit into an engineering schedule.

Capital and skepticism arrived in the same breath. Goldman Sachs published a revised humanoid market forecast raising its 2035 total addressable market estimate sixfold, citing end-to-end AI training and falling unit costs; SoftBank’s Masayoshi Son told CNBC that physical AI is where he expects the next trillion-dollar company to emerge. Yet Texas filings revealed Tesla’s robotaxi fleet there is roughly one-tenth the size of Waymo’s, a reminder that operating scale, not announced ambition, is what distinguishes the leaders in autonomous systems.

Underneath the headlines, the research and tooling base kept widening. The CVPR 2026 ManipArena competition opened a real-robot manipulation benchmark to any team in the world with zero hardware cost. Locus Robotics’ Array system and its Nexera mobile-manipulation acquisition sketched a non-humanoid path to autonomous shelf-to-tote fulfillment. Intuitive Surgical previewed da Vinci 5 upgrades including extended-use Force Feedback instruments, pushing the long-running surgical platform toward what its engineers call the Quintuple Aim.

Items

NVIDIA and Unitree Unveil the Isaac GR00T Reference Humanoid Robot

On June 1 at NVIDIA’s Computex events in Taipei, the company announced the Isaac GR00T Reference Humanoid Robot, the first open humanoid reference design built on NVIDIA’s Jetson Thor compute platform and the Isaac GR00T open development stack. The reference robot uses Unitree Robotics’ newly announced H2 Plus chassis, fitted with Sharpa Wave tactile five-finger hands and powered by the Jetson AGX Thor T5000, which NVIDIA rates at 2,070 FP4 teraflops within a 40-watt envelope.

The strategic point is the openness, not the hardware. NVIDIA packaged the platform with its data collection tool Isaac Teleop, the GR00T family of pre-trained foundation models, the Isaac Sim and Isaac Lab simulation environments, and Isaac ROS deployment middleware. GR00T N1.7 is available in early access with commercial licensing, and the next-generation GR00T N2 was previewed; NVIDIA reports it succeeds at new tasks in new environments more than twice as often as leading vision-language-action baselines.

For academic labs and startups, the calculus changes. Previously, building a research-grade humanoid required a multi-month integration of chassis, sensors, hands, compute, and software. The Isaac GR00T reference, which will be available from Unitree in late 2026, packages all of that into a known-good baseline that the broader ecosystem can extend. Launch partners include Stanford, ETH Zurich, the Allen Institute for AI, and UC San Diego.

The wider effect, if NVIDIA’s bet plays out, is to seed dozens of labs with a comparable platform so that research progress — particularly in vision-language-action models — can be shared, reproduced, and benchmarked rather than tied to bespoke hardware. That standardization is the precondition for the kind of community velocity that vision and language models already enjoy.

Source: NVIDIA Newsroom

Amazon Unveils a Conversational Next-Generation Proteus Robot

Amazon used its Delivering the Future event in London on June 5 to unveil the next generation of its Proteus warehouse robot, part of a wider €10 billion European investment in fulfillment infrastructure. The original Proteus, deployed in 2022 and currently operating in 25 U.S. fulfillment centers, was a quiet workhorse limited to dock areas and capable of transporting carts of up to 400 kilograms. The new version expands the operating envelope across entire fulfillment and delivery sites.

The headline capability is conversational control. Employees can direct the robot using ordinary language — “bring this tote to packing station seven” — rather than a programming interface or technical command syntax. That changes the staffing model: any worker becomes a robot supervisor without specialized training, and the robot’s tasks can be reassigned on the fly as warehouse priorities shift.

Amazon plans to roll the next-generation Proteus into European operations in the first half of 2027. The company also disclosed that it now operates over a million robots across its global fulfillment network — a deployment scale that defines the realistic upper bound of what “robots in production” looks like today.

The deeper signal is about the user interface. As foundation models become reliable enough to translate human intent into structured task graphs, the cost of integrating robots into existing workflows drops sharply. Proteus is one of the first warehouse-scale demonstrations that the natural-language interface is no longer the limiting factor.

Source: About Amazon

Hyundai and Boston Dynamics Send Atlas to “School of Football”

Hyundai launched a five-part episodic film campaign starring Boston Dynamics’ Atlas humanoid, titled “School of Football” and tied to Hyundai’s FIFA World Cup 2026 sponsorship. The centerpiece is a move Hyundai calls the “Ghost Rabona,” a cross-leg strike requiring precise dynamic balance and split-second timing, performed by the actual electric Atlas on a real pitch without CGI.

The technical pipeline is more interesting than the marketing wrapper. Hyundai’s team captured motion data from professional footballers using a motion-capture system, then ran a retargeting process to convert human movements to fit Atlas’s different body structure. The retargeted trajectories became the training signal for a reinforcement learning controller that gave Atlas full-body coordination, dynamic balance, and real-time adaptation to perturbations.

What makes the demonstration substantive is what it implies about industrial deployment. Hyundai is preparing to train Atlas at the Robot Metaplant Application Center within Hyundai Motor Group Metaplant America in Savannah, Georgia, and is already taking delivery of the first 2026 production units. The same machine-learning pipeline that lets Atlas perform a rabona can in principle let it learn factory tasks from human-operator demonstrations.

Hyundai Motor Group and Boston Dynamics have stated plans to build a robotics factory capable of producing 30,000 Atlas robots per year, with operations slated to begin by 2028. Marketing campaigns of this kind are unusual in industrial robotics — but they serve a real purpose, which is to make the case to workers, regulators, and the public that humanoid coworkers are arriving and are designed to be learnable, not magical.

Source: Hyundai Motor Group Newsroom

BYD Confirms Humanoid Robot Program and 20,000-Unit Factory Plan

The Chinese automaker BYD confirmed this week that it is developing humanoid robots, ending years of speculation. Executive Vice President Stella Li disclosed the program, codenamed “Yao-Shun-Yu” and run by BYD’s 15th Business Unit, which focuses on electronic integration and intelligence. The project was actually initiated in 2022, putting BYD on roughly the same timeline as Figure and Tesla’s Optimus program.

Li said roughly 150 prototypes are already being tested inside BYD factories, and that the company intends to deploy 20,000 units in its own production lines this year. A new industrial park in Xi’an is expected to eventually produce 50,000 robots annually. If the household market materializes, BYD has indicated it will distribute through its existing automotive dealer network, leveraging an asset competing humanoid startups cannot match.

Li’s framing of the competitive landscape is striking. She said Chinese robots have not yet developed an adequate “brain,” while U.S. competitors face the opposite problem of underdeveloped “limbs” — the mechanical, actuation, and manufacturing pieces. BYD’s stated goal is parity on both axes, using the company’s vertically integrated battery, motor, and electronics supply chain to build practical machines.

BYD’s entry matters because, like Tesla, it brings both an in-house deployment customer (its own enormous manufacturing footprint) and a capital structure that can absorb years of unprofitable scaling. The implication is that the humanoid market is converging on a small set of vertically integrated industrial conglomerates rather than pure-play startups.

Source: CnEVPost

CVPR 2026 Launches ManipArena, a Real-Robot Manipulation Benchmark

The CVPR 2026 conference, running June 3–7 at the Colorado Convention Center in Denver, hosts the inaugural ManipArena Competition — a real-robot manipulation challenge built around 20 physical-world tasks covering execution reasoning, semantic reasoning, and mobile operation. The competition was jointly initiated by Sun Yat-sen University and Mohamed bin Zayed University of Artificial Intelligence, with X Square Robot as a founding partner.

The structural innovation is the participation model. ManipArena offers zero hardware entry cost: teams submit models that run on remote real-robot infrastructure rather than buying or borrowing arms themselves. Participants get access to 188 hours of verified real-robot operational data for training and validation, along with transparent and interpretable evaluation metrics that go beyond leaderboard ranking.

For a research community that has spent the last decade publishing on simulated manipulation benchmarks of debatable transferability, this is a meaningful change. ManipArena evaluates against actual hardware in actual environments, which means submissions cannot rely on simulator quirks or perception shortcuts. Winning teams present at CVPR, which gives strong work both a public platform and a peer-validated track record.

The broader trend ManipArena represents is the maturation of manipulation from a research curiosity into something measurable. Vision-language-action models have improved so rapidly that the field needs shared benchmarks just to compare like-for-like progress. ManipArena’s emphasis on remote execution may become a template for other expensive-hardware research areas, lowering barriers to entry the way cloud GPU access did for deep learning a decade ago.

Source: CVPR 2026

Goldman Sachs Raises Its 2035 Humanoid Robot Forecast Sixfold

Goldman Sachs Research has revised its 2035 humanoid robot total addressable market estimate from $6 billion to $38 billion, a roughly sixfold increase. The bank’s analysts attribute the revision to two developments that have outpaced their earlier assumptions: end-to-end AI training methods that allow models to learn behaviors without hand-engineered policies, and faster-than-expected cost declines for the underlying robot hardware.

The bank’s manufacturing cost estimates are concrete. A year ago Goldman pegged humanoid unit costs at roughly $50,000 for lower-end models and $250,000 for state-of-the-art versions; the current estimate is $30,000 to $150,000. Goldman’s base case calls for more than 250,000 humanoid robot shipments in 2030, almost entirely for industrial use, climbing to roughly 1.4 million units by 2035.

The investor sentiment surrounding the forecast is unusually emphatic. SoftBank Group CEO Masayoshi Son told CNBC on June 1 that humanoid and industrial robotics, “with physical AI as a core,” is where he expects the next trillion-dollar company to emerge. Coverage from CNBC’s June 3 piece tied Goldman’s revision and Son’s thesis together to describe an investment race entering a more concrete phase.

It is worth holding the optimism in tension with execution risk. Goldman’s prior $6 billion estimate is roughly two years old; the new one rests on assumptions about training efficiency and supply chains that remain unproven at scale. The substantive claim, however, is harder to dispute: end-to-end learned policies and falling hardware costs together change what an investor can reasonably model for the late 2020s and beyond.

Source: Goldman Sachs

Figure’s BotQ Factory Reaches One Humanoid Robot Per Hour

Figure announced that its BotQ manufacturing facility has scaled Figure 03 production from one robot per day to one robot per hour — a 24-fold throughput increase achieved in fewer than 120 days. The company reports having produced more than 350 third-generation humanoid robots so far and manufactured over 9,000 actuators across more than 10 different SKUs.

The throughput milestone reframes the constraint. For most of humanoid robotics’ commercial history the limiting factor has been manufacturing rather than software: a robot capable of useful work was unbuildable at any meaningful price. Figure’s 2026 roadmap targets an initial capacity of 12,000 units per year, scaling to 50,000 annually and a four-year goal of 100,000 robots. The company also intends to fully decouple its supply chain from China by mid-2026.

Practical capability tests have begun catching up. Figure livestreamed a group of Figure 03 robots processing packages nonstop for nearly a week in May, and the company ran a ten-hour competition between a single robot and a human operator in which the robot achieved 98.5% of the human’s throughput. Those numbers are not yet at the level of production deployment, but they no longer require a footnote.

The Figure 03 ramp matters most as evidence of an industrial learning curve. Translating a working prototype into a thousand identical, reliable units is the part of humanoid robotics that has been historically underpriced. Figure’s report suggests the curve is steeper than skeptics expected; whether other entrants can match it will determine the shape of the 2027 market.

Source: Figure

Locus Robotics Acquires Nexera, Adding Patented Mobile Manipulation to Array

Locus Robotics, the long-standing autonomous mobile robot vendor used in hundreds of warehouses, announced on May 19 the acquisition of Nexera Robotics, a mobile-manipulation specialist whose patented technology Locus describes as a breakthrough for autonomous picking. The acquisition reinforces Locus’s Locus Array system, a 10-foot-tall, 1,000-pound autonomous unit that rolls up to warehouse shelving, grabs a plastic tote, lowers it to a platform, and lets an integrated robotic arm pick directly from it.

Array is interesting precisely because it is not a humanoid. While Figure, BYD, Boston Dynamics, and others are building human-shaped general-purpose robots for warehouses, Locus has chosen task-shaped automation: an architecture optimized for shelf-to-tote movement that can automate up to 90% of pick-and-replenish work without trying to walk on two legs. Early deployments are already running with DHL Supply Chain at a site in Columbus, Ohio.

The Nexera acquisition addresses Array’s known weakness — fine manipulation of irregular SKUs. Nexera’s patents are reported to focus on the perception-to-grasp pipeline that lets a stationary arm pick items it has never seen before. Combining that with Locus’s installed base of mobile robots produces a complete autonomous fulfillment loop that does not require a humanoid form factor.

The competitive lesson is that warehouse automation is bifurcating. Humanoids promise generality across the entire facility’s tasks, but for the specific high-volume workflows that dominate fulfillment economics, purpose-built systems like Array remain hard to beat. Both architectures will likely coexist for years, with humanoids filling in the long tail of irregular tasks that bespoke systems cannot economically address.

Source: Locus Robotics

Texas Filings Show Tesla Robotaxi Fleet One-Tenth the Size of Waymo’s

Public filings analyzed by CNBC on May 28 reveal that Tesla’s robotaxi service has 42 automated vehicles registered in Texas, against Waymo’s 577 — roughly one-tenth the size. Tesla also trails AV Ride, a smaller player with 317 authorized vehicles, and is only just ahead of Amazon-owned Zoox at 35. The data is significant because Tesla has positioned its Cybercab and Full Self-Driving program as imminently scalable, while Waymo has built incrementally toward today’s operating footprint over a decade.

The Texas numbers reflect a broader pattern. Waymo currently operates revenue-generating robotaxi service in Phoenix, San Francisco, Los Angeles, Austin, and Atlanta, with public expansions planned for Miami, Dallas, Houston, San Antonio, Orlando, and Denver. Waymo’s grown fleet has reached approximately 1,000 vehicles globally and the company reports millions of paid rides per month in San Francisco alone.

Zoox plans to launch a paid robotaxi service in San Francisco in the second half of 2026 using its purpose-built, no-steering-wheel vehicle platform. The Zoox approach is architecturally distinct from Waymo’s retrofitted Jaguar I-Pace SUVs and is the most direct test of whether vehicles designed from scratch for autonomy will outperform adapted production cars at scale.

The headline is not that Tesla is failing — the company can scale rapidly once its software clears regulatory gates — but that operating reality continues to diverge from public expectations. The robotaxi market is being shaped now not by future promises but by who is moving real passengers on real roads, with real safety records that regulators can audit.

Source: CNBC

Intuitive Surgical Previews da Vinci 5 Force Feedback and Telepresence Upgrades

Intuitive Surgical, the dominant incumbent in soft-tissue robotic surgery, previewed a series of upgrades to its da Vinci 5 platform on May 21. The upgrades span telepresence, mobile login for surgeons, an updated SimNow 2 training environment, extended-use Force Feedback instruments, new intraoperative tools including a digital ruler, and enhanced encryption for security and reliability — packaged under Intuitive’s framing of the “Quintuple Aim” of healthcare quality, equity, efficiency, clinician experience, and patient experience.

The Force Feedback instrument extension is the substantive clinical change. Five of six Force Feedback instruments are being cleared to increase from six to fifteen uses, with one cleared for ten uses. That cuts the per-procedure consumables cost of using haptic instruments and supports continued evidence generation on how restoring force-sensing to the surgeon improves outcomes — a question da Vinci platforms historically had to leave unanswered because the earlier generations did not transmit force.

Telepresence and mobile login matter to a different audience. The ability for a senior surgeon to log in remotely to assist or supervise a procedure underway in another hospital extends the reach of the small population of specialists trained in the most complex robotic procedures, which has been a stated goal of surgical robotics since the field’s inception.

The broader market context is that Intuitive’s incumbency is being squeezed from both ends. Multinationals including Medtronic with the Hugo system and Johnson & Johnson’s pending Ottava platform are challenging in soft tissue, while specialty entrants like CMR Surgical’s Versius Plus and SS Innovations target specific use cases. The da Vinci 5 upgrades suggest Intuitive intends to defend its position through deeper clinical capability rather than price, betting that demonstrated surgical outcomes will keep the platform sticky.

Source: Intuitive Surgical