Humanoid Robotics Weekly — Week of Mar 10 – 16, 2026

Top 3 Headlines

1. NVIDIA GTC 2026 Opens with a Physical AI Blitz — Rubin, GR00T N2, and “Chips the World Has Never Seen”

Jensen Huang took the stage at SAP Center in San Jose on March 16 for the most robotics-heavy GTC keynote in NVIDIA’s history. Over 30,000 attendees — and every major humanoid company — watched as NVIDIA unveiled the Rubin GPU platform (promising 10x lower inference costs vs. Blackwell), a next-gen GR00T foundation model with enhanced multi-embodiment capabilities, and the Newton physics engine now integrated into Isaac Lab for production sim-to-real transfer. The demo reel featured 1X’s NEO autonomously tidying a home using a GR00T-trained policy, and Disney Research’s BDX droids performing expressive movements on stage. With 240+ Inception startups showcasing Physical AI, GTC has become the robotics industry’s de facto annual summit. The message was unmistakable: NVIDIA is building the full-stack operating system for the physical world — from silicon to simulation to software.

2. Physical Intelligence Unveils MEM — Giving Robot Brains Long-Term Memory

Physical Intelligence (Pi), now valued at $5.6B, published Multi-Scale Embodied Memory (MEM) — a breakthrough that gives vision-language-action (VLA) models both short-term and long-term memory. Previously, robot AI operated on single observations or very short histories. MEM solves this with a dual-scale architecture: dense visual memory for fine-grained spatial awareness, and language-based memory for context spanning 15+ minutes. Demonstrations showed robots completing full kitchen cleanups and multi-step cooking tasks from scratch — tasks that require remembering which areas have been cleaned, which ingredients have been used, and what step comes next. This is the first time a robot foundation model has demonstrated “job-scale” memory rather than atomic actions. If this scales, it changes what humanoids can actually do in real-world settings.

3. Tesla Begins Optimus Gen 3 Pilot Production — Fremont Conversion Accelerates

Tesla confirmed that Optimus Gen 3 pilot production is now running at the Fremont factory, where Model S and Model X production lines are being permanently converted to humanoid robot manufacturing. Gen 3 features 50-actuator hands (25 per forearm — a 4.5x increase from Gen 2), giving Optimus the dexterity to handle delicate objects, fold laundry, and catch thrown items. The long-term target: 1 million units per year at $20,000 COGS per robot. Musk was candid that current robots are still “for learning, not productive work,” but the factory commitment is real — Tesla is spending $20B+ in 2026 CapEx on robotics, autonomy, and energy. This is the largest single manufacturing bet on humanoid robots to date.

Company-by-Company Updates

NVIDIA

GTC 2026 keynote (March 16) introduced the Rubin GPU platform — successor to Blackwell — with 10x lower inference costs and 4x fewer GPUs for mixture-of-experts models. Meta announced a deal to buy millions of Vera Rubin units.
Newton physics engine (co-developed with Google DeepMind and Disney Research) now available in Isaac Lab for production-grade sim-to-real transfer. MuJoCo-Warp integration accelerates robotics ML workloads by 70x.
GR00T ecosystem expanding: 1X, Agility Robotics, Boston Dynamics, Apptronik, Figure AI, Fourier, Sanctuary AI, Unitree, and XPENG all building on NVIDIA's platform. Over 240 Inception startups showcased Physical AI at GTC.
Skild AI CEO Deepak Pathak featured in GTC pre-keynote alongside Perplexity, LangChain, and Mistral AI leaders — signaling NVIDIA's embrace of the 'robot brain' startup ecosystem.
Jensen Huang teased 'several new chips the world has never seen before' — with Feynman architecture (2028) breadcrumbs surfacing in breakout sessions.

Physical Intelligence

Published MEM (Multi-Scale Embodied Memory) paper with Stanford, UC Berkeley, and MIT. First robot foundation model to handle 15-minute task horizons with both visual and language-based memory.
MEM solves 'causal confusion' — the phenomenon where naive memory systems actually harm robot performance by amplifying spurious correlations in imitation learning.
Published 'The Physical Intelligence Layer' manifesto (Feb 24) arguing general-purpose physical intelligence models will trigger a 'Cambrian explosion' of robotics applications.
Total funding: $1.07B across Series A ($400M) and Series B ($600M). Investors: Thrive Capital, Lux Capital, CapitalG, Khosla, Sequoia. Valuation: $5.6B.

Tesla Optimus

Gen 3 pilot production line running at Fremont. 'Significantly larger Gen 3 production line coming in 2026.' Model S/X end-of-production confirmed for Q2 2026 — those lines go to Optimus.
Gen 3 hands: 25 actuators per forearm/hand (50 total), a 4.5x increase over Gen 2. Demonstrations show paper-tearing, cabinet-opening, laundry-folding, and catching thrown objects.
1,000+ Optimus robots deployed across Tesla factories for data collection. Tasks: battery cell sorting, component moving, kitting, visual inspection. Grok AI integrated for voice interaction.
Target timeline: mid-2026 low-volume production (thousands of units for internal use), late 2026 significant production, late 2026 limited external sales to select partners (~$30K initial price). Consumer sales: end of 2027.
Giga Texas expansion continuing: dedicated humanoid factory under construction alongside Fremont conversion. Combined target: 1M units/year at scale.

Figure AI

BotQ manufacturing facility fully operational — first-gen line capable of 12,000 humanoids per year. Vertical integration: Figure's own robots help assemble key components on the line.
Figure 03 specs locked: 5'6" tall, 61 kg, 20 kg payload, 5-hour battery, wireless inductive charging at 2 kW via foot pads. Fingertip sensors detect forces as small as 3 grams.
Helix AI advancing: VLA model learns tasks from demonstration videos — achieved towel-folding capability with only 80 hours of training footage.
Series C: $1B+ at $39B valuation. Investors: Parkway Venture Capital, Brookfield, NVIDIA, Macquarie Capital, Intel Capital. Target: 100,000 robots over next four years.
Home alpha testing with select partners underway. Broader home availability targeted for late 2026 — though analysts remain cautious on the timeline.

Agility Robotics

Toyota Motor Manufacturing Canada commercial deployment expanding: seven Digit humanoids now deploying to Woodstock, Ontario RAV4 plant under Robots-as-a-Service (RaaS) agreement. First commercial humanoid deployment in Canadian automotive manufacturing.
RoboFab capacity scaling from hundreds toward thousands of Digit units. Peak capacity: up to 10,000 units annually.
CEO Peggy Johnson: 'With our next generation of Digit, we will be the first company to deliver the first cooperatively safe humanoid robot to work alongside people.' Safety certification push underway.
Fortune 500 customer roster now includes: GXO, Schaeffler, Amazon, Toyota, Mercado Libre. 80% of Digit's ~5,000 parts U.S.-sourced.

1X Technologies

NEO featured prominently in Jensen Huang's GTC keynote — demonstrated autonomously performing domestic tidying tasks using a GR00T-trained policy from the 1X/NVIDIA AI collaboration.
World Model (January 2026) enabling NEO to teach itself new tasks from voice/text prompts using video-based visualization. Demos: packing lunchboxes, operating doors, ironing shirts — all without prior training examples.
NEO specs: 168 cm, 30 kg, 70 kg lift, 22-DoF hands (44 total), 842 Wh battery (~4 hours), NVIDIA Jetson Thor compute. Operating noise: 22 dB (below a refrigerator).
EQT partnership: up to 10,000 NEO units by 2030 across 300+ portfolio companies in manufacturing, warehousing, and logistics.
U.S. deliveries on track for 2026 — hybrid autonomy model with human teleoperator backup. Privacy controls: room-level opt-out, audio masking, visual blurring.

Apptronik

Total funding: $935M at $5.3B valuation. Google co-led the $520M Series A extension. Using capital to expand Austin footprint and open California office.
Apollo robots in active testing at Mercedes-Benz and GXO Logistics factories. Mercedes planning expansion beyond German and Hungarian plants if pilot KPIs hold.
Google DeepMind Gemini Robotics integration deepening — multi-embodiment control enabling same AI to operate dual-arm industrial robots and full humanoids.
CEO Jeff Cardenas targeting $1B in robot orders starting 2027 at ~$80K/year per unit. Next-gen Apollo debut expected before Q2.

Boston Dynamics

All 2026 Atlas production units remain fully committed to Hyundai RMAC and Google DeepMind. No new customer availability until 2027.
Production-ready Atlas specs: 6.2 feet tall, 7.5-foot reach, 66 lb payload, 56 degrees of freedom. Limbs replaceable in under 5 minutes. Autonomous battery swap for continuous operation.
CTO Pras Velagapudi presenting at GTC 2026 — among featured robotics speakers alongside Agility, Physical Intelligence, and Tesla representatives.
Hyundai Mobis supplying actuators for Atlas. 30,000-unit/year factory in planning. Metaplant America (Savannah, GA) deployment on track for 2028.

SoftBank / ABB Robotics

SoftBank's $5.4B acquisition of ABB Robotics progressing toward mid-2026 close. Integration planning underway with existing portfolio: Berkshire Grey, Agile Robots, Skild AI ($14B), AutoStore — all under 'Robo HD' holding.
ABB Robotics: 7,000 employees, $2.3B revenue, serving BMW and major automotive/electronics manufacturers worldwide. This gives SoftBank instant industrial robotics scale.
Masayoshi Son's 'Physical AI' empire taking shape — from Skild AI's universal robot brain to ABB's industrial arms to Fourier's humanoids. No other investor has this breadth of robotics portfolio.

Chinese Ecosystem

Xpeng breaking ground on first 'full-chain' humanoid robot mass production base in Guangzhou (110,000 sq meters). Next-gen IRON: all-solid-state batteries, 2,250 TOPS AI compute (three Turing chips), VLA 2.0 brain. Mass delivery targeted end of 2026.
UBTECH production capacity at 300 units/month. Targeting 5,000 humanoid deliveries in 2026, 10,000 by 2027. Cumulative Walker series orders: 1.3 billion yuan. Factory deployments: Geely, FAW-Volkswagen, BYD, Audi FAW, BAIC, Foxconn.
Unitree preparing for Shanghai STAR Market IPO at ~$7B valuation. Shipped 5,500+ humanoids in 2025 — 36x more than U.S. rivals Figure and Tesla. Targeting 20,000 units in 2026. Spring Festival Gala showcase featured acrobatic wushu techniques.
U.S. government scrutiny increasing: House Select Committee requesting investigations into Unitree over alleged PLA connections. Security researchers flagging data collection concerns. Western agencies demanding penetration tests before approving new humanoid shipments.

Patent Watch

The patent landscape continues to bifurcate between hardware incumbents and AI-native challengers. LG Electronics maintains its position as the top U.S. patent holder in broad robotics with 1,000+ patents. The fastest-growing filing categories: VLA architectures, sim-to-real transfer methods, and multi-embodiment control systems.

NVIDIA — Rubin Platform + Newton Physics Engine IP

GTC 2026 revealed the depth of NVIDIA's Physical AI IP stack. The Rubin platform (successor to Blackwell) combines Vera CPU + Rubin GPU for robotics inference at 10x lower cost. Newton — co-developed with Google DeepMind — is an open-source physics engine optimized for robot learning, with MuJoCo-Warp delivering 70x acceleration on ML workloads. NVIDIA is patenting the full pipeline: from synthetic data generation (Cosmos) through simulation (Isaac/Newton) to on-robot inference (Jetson Thor). This is the most comprehensive Physical AI IP stack being assembled by any company.

Physical Intelligence — Multi-Scale Embodied Memory (MEM)

Pi's MEM paper introduces a novel dual-scale memory architecture for VLA models — dense visual memory for spatial awareness + language-based memory for long-horizon context. The approach solves 'causal confusion' in imitation learning and enables 15-minute task horizons. This is foundational IP for any company building robots that need to remember what they've done and plan what comes next.

Flexion Robotics — Horizontal Autonomy Stack

Zurich-based Flexion ($50M Series A from DST Global, NVentures, Prosus) is building the 'Windows of humanoid robotics' — a three-layer autonomy stack: Command Layer (LLMs), Motion Layer (synthetic data-trained), Control Layer (modular skill library). Founded by ex-NVIDIA researchers from ETH Zurich. If the horizontal model wins, Flexion's IP becomes the middleware layer between NVIDIA's silicon and every hardware maker.

Tesla — Optimus Gen 3 Actuator Architecture

Gen 3 hands (patent WO2024073135A1 and related filings) feature 25 actuators per forearm/hand — a system for real-time trajectory smoothing and coordinated motion across 50 total hand actuators. Tesla's patent strategy focuses on manufacturing-friendly actuator designs compatible with automotive supply chains, giving them a production cost advantage that pure robotics companies lack.

⚠ The “Robot Brain” IP Race Intensifies

Five companies are now competing to own the intelligence layer for humanoid robots: NVIDIA (GR00T + full stack), Physical Intelligence (MEM + cross-embodiment), Skild AI ($14B, universal OS), Flexion ($50M, horizontal autonomy), and Google DeepMind (Gemini Robotics). Each is filing aggressively around VLA architectures, memory systems, and sim-to-real transfer. Meanwhile, the Green-VLA paper (arXiv, Jan 2026) demonstrates that a single staged VLA policy can control humanoids, mobile manipulators, and fixed-base arms — suggesting the “one brain, any body” paradigm is technically achievable. The IP winner here captures the platform economics of the entire industry.

Hiring Signals

GTC week amplifies hiring signals across the sector. The NVIDIA effect is visible: companies building on NVIDIA’s platform are hiring for integration roles, while companies building their own stacks are hiring for foundational AI research.

Company	Notable Postings	What It Tells You
Tesla	110+ Optimus roles persist; new Gen 3 production line engineers, Fremont conversion specialists	The factory conversion is real and accelerating. This isn't a concept — it's a manufacturing commitment with headcount to match.
NVIDIA	2026 PhD Robotics Research internships; Isaac/GR00T platform engineers; Physical AI developer relations	Building the ecosystem layer — developer relations hires signal NVIDIA wants GR00T adoption to be self-sustaining, not just demo-driven.
Google DeepMind	Research Scientists for Robotics team — VLA safety, HRI scenarios, physical safety with foundation models	The safety angle is new. DeepMind is hiring for 'safe humanoid robot actions' — a prerequisite for moving Atlas beyond controlled factory environments.
Figure AI	"Humanoid Robot Pilot — Helix Team" (San Francisco); BotQ production engineers	The 'Pilot' title means real-world home testing is active. Combined with production hiring, Figure is running deployment and manufacturing in parallel.
Agility	RoboFab production scale-up roles; Toyota deployment support engineers; next-gen Digit cooperative safety engineers	Scaling from hundreds to thousands of units. 'Cooperative safety' hires = preparing Digit to work alongside humans, not in isolated zones.
Physical Intelligence	Foundation model researchers; robotics data infrastructure engineers	Post-MEM, Pi is scaling its data pipeline — the bottleneck for long-horizon robot intelligence. Expect more memory-augmented demos soon.

The macro signal: Two distinct hiring waves are now visible. Wave 1: production and deployment roles (Tesla, Agility, Figure) for companies shipping hardware. Wave 2: AI research and platform engineering roles (NVIDIA, Physical Intelligence, DeepMind) for companies building the intelligence layer. The sector is splitting into hardware shippers and brain builders — and both are hiring aggressively.

Funding & Deals

SoftBank acquires ABB Robotics — $5.4B

The largest robotics deal of the decade. SoftBank absorbs 7,000 employees and $2.3B in revenue, integrating with Skild AI, Berkshire Grey, Agile Robots, and AutoStore under 'Robo HD.' Masayoshi Son is assembling the most comprehensive robotics portfolio in the world — from foundational AI to industrial arms to humanoid bodies.

Flexion Robotics — $50M Series A

ETH Zurich spinout building the 'horizontal autonomy stack' for humanoid robots. DST Global Partners, NVentures (NVIDIA's venture arm), Prosus, Redalpine, and Moonfire participated. Total raised: $59.3M. The NVIDIA investment signals alignment with the GR00T ecosystem strategy.

Agility Robotics — Toyota Canada RaaS deal

Seven Digit humanoids deployed to TMMC's Woodstock RAV4 plant under Robots-as-a-Service agreement. First commercial humanoid deployment in Canadian automotive. The RaaS model lowers adoption barriers — Toyota pays subscription, not upfront capital.

Xpeng Robotics — Guangzhou factory groundbreaking

110,000 sq meter 'full-chain' mass production base breaking ground Q1 2026. First humanoid robot factory purpose-built by an EV company. Mass delivery of IRON humanoid targeted end of 2026.

Physical Intelligence — $1.07B total raised

Series B ($600M) closed late 2025 at $5.6B valuation. Investors: Thrive Capital, Lux Capital, CapitalG. MEM paper validates the investment thesis — Pi is building the foundational memory and reasoning layer for all robot AI.

Meta — Massive Vera Rubin GPU order

Meta announced a deal to buy millions of NVIDIA Vera Rubin units + Grace CPUs for AI infrastructure. While not robotics-specific, this order funds the R&D behind every NVIDIA robotics product — and signals that Physical AI compute demand is being subsidized by hyperscaler AI spending.

Total robotics / Physical AI investment in 2026 so far: ~$16B+
SoftBank’s ABB acquisition alone adds $5.4B to the 2026 total. Combined with VC activity ($14B+ in Jan–Feb), the sector has absorbed more capital in 11 weeks than in any full year in robotics history. GTC week is expected to trigger another wave of announcements.

One Insight: The “Memory Wall” Is the New Frontier

This week revealed the next critical bottleneck in humanoid robotics — and it isn’t hardware, locomotion, or even general dexterity. It’s memory.

Every humanoid robot today operates in a perpetual present. Current VLA models process a single image or a few seconds of history, then output an action. They can fold a towel or pick up a cup. But ask them to clean an entire kitchen — remembering which counters have been wiped, which dishes have been washed, which cabinet was already organized — and they fail. The robot doesn’t remember what it just did 30 seconds ago.

Physical Intelligence’s MEM changes this equation. By giving robots both short-term spatial memory (dense visual data for fine manipulation) and long-term semantic memory (language-based summaries for multi-minute context), MEM enables what Pi calls “job-scale” tasks — the kind of work that lasts 15 minutes, not 15 seconds. The demos are striking: a robot cooking grilled cheese from scratch, following a recipe, adapting when ingredients aren’t where expected.

This matters for every company in the space. NVIDIA’s GR00T, Skild’s Brain, Flexion’s autonomy stack, and Figure’s Helix will all need to solve the memory problem to move beyond atomic tasks. The Green-VLA paper (Jan 2026) showed one policy can control any robot body. MEM shows one policy can remember what it’s doing across an entire job. Put them together and you have the technical foundation for a robot that can actually replace a shift worker — not just perform a single demo.

The hardware race gets the headlines. The brain wars get the VC dollars. But the memory wall is what separates a $20,000-a-year demo robot from a $200,000-a-year productive worker. The companies that solve memory at scale — not just in a lab — will define which humanoids actually get deployed in factories, warehouses, and homes. Watch for every major AI robotics lab to publish memory architectures in the next 90 days.

Looking Ahead

NVIDIA GTC 2026 continues (March 17–19): Breakout sessions on GR00T, Newton, and Physical AI. Investor Q&A on March 17. All-In Podcast live from show floor on March 18.

Apptronik next-gen Apollo: CEO Jeff Cardenas says debut is imminent — watch for a reveal during or immediately after GTC week.

Unitree IPO watch: Shanghai STAR Market listing expected mid-2026 at ~$7B. Any prospectus filings in March would be significant.

Tesla Optimus Gen 3 reveal: Full production-intent prototype expected this quarter. Fremont conversion progress will be visible in Q1 earnings commentary.

SoftBank / ABB Robotics: Regulatory approval process underway. Mid-2026 close expected. Integration planning with Skild AI and Robo HD portfolio will shape the competitive landscape.