Skild AI
Last reviewed
Jun 5, 2026
Sources
41 citations
Review status
Source-backed
Revision
v4 ยท 7,546 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
Jun 5, 2026
Sources
41 citations
Review status
Source-backed
Revision
v4 ยท 7,546 words
Add missing citations, update stale details, or suggest a clearer explanation.
Skild AI is a robotics artificial intelligence company building a general-purpose foundation model for physical embodiments, described by the company as a single "robot brain" capable of controlling many different robot bodies across many different tasks. The startup was co-founded in May 2023 by Carnegie Mellon University professors Deepak Pathak (chief executive officer) and Abhinav Gupta (president), both researchers with extensive backgrounds in self-supervised learning, computer vision, and robot learning.[27] Its central product is the Skild Brain, a robot foundation model designed to be "omni-bodied": a single trained model that runs across humanoids, quadrupeds, mobile manipulators, and tabletop arms without a body-specific controller for each platform.[3]
Skild AI emerged from stealth in July 2024 with a $300 million Series A round at a $1.5 billion valuation, widely covered at the time as one of the largest Series A rounds in robotics history.[1][8] By January 2026 the company had raised a further $1.4 billion at a valuation of more than $14 billion in a Series C led by SoftBank Group with NVentures (NVIDIA's venture arm).[2][10] In April 2026 it acquired Zebra Technologies' robotics automation business (formerly Fetch Robotics) along with the Symmetry Fulfillment orchestration platform.[6][11] Skild AI is headquartered in Pittsburgh with offices in the San Francisco Bay Area and, since February 2026, Bengaluru, India.[5]
Skild AI sits within the small group of robotics companies that take a software-first or "horizontal" approach: they build the AI model that controls a robot, but they do not primarily build the robot hardware itself. The thesis is that intelligence is the bottleneck in real-world robotics, and that a single large model trained across many bodies and tasks will outperform narrow controllers handcrafted per platform.[3]
The company's founders frame their goal as "any robot, any task, one brain." The technical approach is closer to the playbook used to train large language models than to classical robotics: assemble a large mixed corpus from large-scale physics simulation, internet videos of humans, teleoperation data, and real-world robot data, then train a single transformer-based model that can be conditioned on different morphologies.[4] That orientation places Skild AI alongside companies such as Physical Intelligence, NVIDIA's Isaac GR00T, and Google DeepMind's RT-2 work, and distinguishes it from full-stack humanoid companies such as Figure AI, 1X Technologies, and Tesla, which build both the robot and the model in-house.
At the World Economic Forum in Davos in January 2026, CEO Deepak Pathak argued that physical intelligence, rather than language models, is the genuine bedrock of artificial general intelligence, and challenged conventional wisdom that self-driving datasets from companies like Tesla would transfer meaningfully to general-purpose manipulation robots.[30] He framed the company's mission in terms of closing what he calls the "Moravec gap": the long-standing observation that things easy for humans (folding a towel, opening a door, climbing stairs) remain extremely hard for machines, while the symbolic tasks once thought to define intelligence (chess, arithmetic, summarising text) turned out to be relatively easy.[30] Pathak's argument is that the recent generative AI boom solved the easy half of the gap and that the harder half can be closed only by training on physical interaction data at a scale not previously attempted.[30]
Deepak Pathak is the chief executive officer. He was the Raj Reddy Associate Professor at Carnegie Mellon University's School of Computer Science before going on leave to start Skild AI, with appointments in the Robotics Institute and the Machine Learning Department.[33] He earned his PhD at UC Berkeley in 2019 under Alyosha Efros and Trevor Darrell, and his bachelor's at IIT Kanpur, where he received a gold medal.[33] Before CMU he was a researcher at Meta AI Research with Jitendra Malik, and a visiting postdoctoral researcher at Berkeley with Pieter Abbeel.[33] His honors include the Sloan Research Fellowship (2025), the MIT Technology Review TR 35 Under 35 Innovator Award (2024), the Raj Reddy Chair at CMU (2024), and the Okawa Research Award (2022).[33] Best-known papers: "Curiosity-driven Exploration by Self-supervised Prediction" (ICML 2017) and "Learning to Control Self-Assembling Morphologies" (NeurIPS 2019).[35][36]
Abhinav Gupta is the president and co-founder. He was a tenured professor at the Carnegie Mellon Robotics Institute before going on leave for the company. He earned his PhD at the University of Maryland in 2009 and his bachelor's at IIT Kanpur in 2004, arrived at the CMU Robotics Institute as a postdoc in 2009, and joined the faculty in 2011. From 2018 he co-led a new Facebook AI Research robotics lab in Pittsburgh, working part-time with Meta while keeping his CMU duties, and returned to full-time CMU work in May 2022.[34] His research has focused on self-supervised visual representation learning, large-scale visual learning, and reasoning, with awards including the Sloan Research Fellowship, the ONR Young Investigator Award, the PAMI Young Researcher Award, and the Bosch Young Faculty Fellowship. By the time of Skild AI's Series A, he had over 75,000 academic citations.
Pathak and Gupta were already research collaborators in Pittsburgh and had discussed launching a company together for nearly a decade before deciding the timing was right in early 2023.[27] Sequoia partner Stephanie Zhan, who led one of the Series A checks, has described the founding pitch as one she heard "once in a blue moon."[21] In a follow-up essay for Sequoia, Zhan articulated the firm's thesis as "a GPT-3 moment is coming to the world of robotics," arguing that the same scaling dynamics that produced GPT-3 from earlier language model research would soon transfer to embodied intelligence, with Skild positioned to play the role OpenAI played in the digital domain.[21]
The team includes other CMU robot-learning researchers and engineers from Meta, Tesla, Nvidia, Amazon, Google, Stanford, UC Berkeley, and UIUC. The company is headquartered in Pittsburgh (registered at 141 South Saint Clair Street), with additional offices in the San Francisco Bay Area and, from February 2026, Bengaluru, India.[5] The Bengaluru opening was Skild's first office outside the United States and is focused on hiring engineers and scientists for perception, control policies, evaluation, and deployment infrastructure.[5]
Through 2025 and the first half of 2026 the company has been an aggressive hirer, with publicly listed roles spanning research scientists in reinforcement learning and locomotion, robotics engineers focused on deployment and testing, perception engineers, infrastructure engineers building simulation pipelines, and operations staff to support customer rollouts. Job postings emphasise hands-on experience with sim-to-real transfer, large-scale distributed training, and field deployment of legged or manipulation systems. The pay band quoted publicly by recruiting trackers runs from roughly $45 per hour for contract deployment engineers to total compensation packages well above $500,000 for senior research staff, reflecting the broader compensation arms race in embodied AI research.
Skild AI is one of the most heavily funded software-only robotics startups. The company has raised at least $1.835 billion across publicly disclosed rounds, with valuation rising roughly tenfold between its Series A and its Series C.
| Round | Date | Amount | Valuation | Lead investor(s) | Other investors |
|---|---|---|---|---|---|
| Stealth / pre-seed | 2023 | Not publicly disclosed | Not disclosed | Not disclosed | Backers later confirmed in Series A list |
| Series A | July 2024 | $300 million | $1.5 billion (post-money) | Lightspeed Venture Partners, Coatue Management, SoftBank Group, Bezos Expeditions | Felicis Ventures, Sequoia Capital, Menlo Ventures, General Catalyst, CRV, SV Angel, Carnegie Mellon University, Amazon Industrial Innovation Fund, Alexa Fund |
| Series B | April 2025 | ~$135 million (SEC-reported) / reported as ~$500 million | $4.5 billion to $4.7 billion | SoftBank Group | NVentures (NVIDIA), Bezos Expeditions, Samsung, LG Technology Ventures, Schneider Electric, CommonSpirit Health, Salesforce Ventures, Disruptive, 1789 Capital |
| Series C | January 14, 2026 | $1.4 billion | More than $14 billion (post-money) | SoftBank Group, with NVentures (NVIDIA) | Macquarie Capital, Bezos Expeditions, Lightspeed, Felicis, Coatue, Sequoia (follow-on); strategic: Samsung, LG, Schneider Electric, CommonSpirit, Salesforce Ventures; new: Disruptive, 1789 Capital, IQT, TF Capital, Andra Capital, Palo Alto Growth Capital, Alpha Square, Mirae Asset, Destiny |
The Series A was pre-circulated in the press in late April 2024 by The Information before being officially announced on July 9, 2024.[1][12] Carnegie Mellon University itself participated as an investor, which is unusual for a US research university.[1] In early 2025 SoftBank entered talks to lead a follow-on round.[13][28] The Series B closed in April 2025 at approximately $4.5 billion valuation. Technical.ly's reporting noted a discrepancy between public claims of $500 million raised and SEC Form D filings showing only about $135 million under Skild-related entities, suggesting some tranches may be structured through vehicles that file separately or not at all under the "Skild" name.[24] By December 2025 SoftBank and NVIDIA were negotiating a larger Series C, which closed on January 14, 2026 at $1.4 billion and a post-money valuation north of $14 billion.[14][15][16] The same announcement disclosed that revenue had grown from zero to roughly $30 million during 2025.[2][10]
The Series C is notable for the breadth of strategic capital it pulled in. Macquarie Capital, IQT (the venture arm associated with the US intelligence community), Mirae Asset of South Korea, and Andra Capital all participated alongside the existing financial backers.[2][18] Crunchbase News flagged the round as one of the fastest valuation jumps in the recent robotics cycle, with the company tripling its valuation in roughly seven months between the Series B close and the Series C announcement.[29]
Skild AI describes its mission as building "intelligence that is grounded in the physical world," sometimes phrased as physical artificial general intelligence.[4] The argument is that current generative AI systems are trained on text and images and have no real grasp of contact dynamics, mass, friction, or affordances; embodied intelligence has to be trained on physical interaction data, not just web scrape.[3]
The technical strategy has several distinguishing pieces:
This is a deliberate contrast with vision-language-action systems such as RT-2 or PaLM-E (see PaLM-E), which the Skild team has criticised for relying on a thin layer of robot data on top of a large vision-language model trained mainly on web content.[3] Skild's view is that physical common sense has to be learned from physical data, not from semantic knowledge alone.[3]
Skild's go-to-market follows a deliberate progression through environments ranked by structural complexity. The company is first deploying in semi-structured settings such as factories, warehouses, and construction sites. The data and experience gathered there is intended to enable deployment in less structured environments such as hospitals and hotels. The long-term goal is deployment in fully unstructured environments, including consumer homes.[4] Pathak has described this progression as building a self-reinforcing data flywheel: each deployed robot streams real-world performance data back into the training loop, making the model progressively more capable, which enables deployment in harder environments.[30]
The data flywheel concept is central to the company's commercial pitch and is what Skild has used to justify accepting industrial deployments that look modest in scale on their own terms. In Pathak's framing, the value of a Foxconn line or a warehouse fleet is not the per-robot revenue but the continuous stream of real-world interaction data each deployed unit feeds back into the training pipeline.[30] That data improves the next generation of the Skild Brain, which in turn unlocks harder environments and more customers. The logic is closely analogous to the way Tesla has argued for the value of its driving fleet for Full Self-Driving data, and to the way OpenAI has framed ChatGPT usage as a training signal for future models. Skild AI's particular wrinkle is that, because the same brain runs across many bodies and tasks, every robot in the field contributes data that helps every other robot, regardless of whether they share a manufacturer.[30]
Skild AI's technology is built on a long research track record from the Pathak and Gupta labs at CMU and Berkeley, much of it published well before the company was founded. "Curiosity-driven Exploration by Self-supervised Prediction" (Pathak et al., ICML 2017) introduced an intrinsic curiosity module that rewards an agent for visiting states whose consequences it cannot yet predict, and is one of the most cited approaches to exploration in deep reinforcement learning.[35] "Learning to Control Self-Assembling Morphologies" (Pathak et al., NeurIPS 2019) studied how primitive limb agents can dynamically link up into composite bodies, using modularity as a route to generalisation; the framing of "one policy, many bodies" is a direct precursor to the omni-bodied brain Skild AI is now building.[36] Gupta's lab spent more than a decade building self-supervised systems that learn visual representations from large unlabeled image and video corpora, including work on visual common sense, action recognition, and large-scale grasping datasets, and earlier collaboration on resources like Visual Genome. More recent papers from the Pathak group on legged locomotion and dexterous manipulation, including rapid motor adaptation, in-the-wild walking on a Unitree quadruped, and dexterous in-hand manipulation, demonstrated that learning-based controllers could ride out the messy real world.
Publicly disclosed details of the Skild Brain remain limited; the company has not published a formal model card or technical report. From conference talks, blog posts, the NVIDIA case study, and press coverage of the July 2025 unveiling, the system is best understood as a unified end-to-end control stack with three tiers:[22][9]
| Tier | Frequency | Role | Inputs | Outputs |
|---|---|---|---|---|
| High-level reasoning policy | Low frequency (sub-Hz to a few Hz) | Task and scene understanding, planning, sequence selection | Camera images, language goal, scene memory | Subgoal directives passed to lower tiers |
| Mid-level manipulation/navigation policy | Mid frequency (tens of Hz) | Mode switching between locomotion, grasping, and other primitives | Subgoal, proprioception, vision | Continuous trajectory targets |
| Low-level control policy | High frequency (hundreds of Hz to kHz) | Joint angles, motor torques, balance | Trajectory targets, proprioception | Direct actuator commands |
The model is conditioned on body identity through proprioceptive feedback rather than through an explicit hardware descriptor, which is part of why the company emphasises that the brain can take over an unfamiliar robot without a body-specific configuration step.[3] Skild's published demonstrations claim that when the brain is started on a quadruped that happens to be standing upright, the policy treats the platform as a bipedal humanoid until contact with the ground tells it otherwise, an example of what the team calls in-context body identification.[9]
For data, the training mixture is a deliberate blend designed to expose the model to physical regularities at scale while keeping the cost of real-world data collection bounded. The dominant source is procedurally generated simulation in NVIDIA Isaac Lab and the Omniverse stack, with thousands of robot instances running in parallel across randomised environments.[22] A smaller share is teleoperation data collected on Skild's own platforms and partner hardware. Internet video of humans performing manipulation and locomotion tasks is used as a weak supervision signal for affordance learning and for grounding language to actions.[4] The company has separately described an experimental project to teach robots to cook by having them watch instructional cooking videos on YouTube, which is being used as a stress test for the affordance learning pipeline rather than as a near-term commercial product.[40]
For training infrastructure, Skild AI uses a combination of on-premise and private cloud compute. In March 2025 the company announced a partnership with Hewlett Packard Enterprise (HPE) to deploy a secure private AI-as-a-service environment built on HPE Cray XD670 servers powered by NVIDIA HGX H200 GPUs for large model training and HPE ProLiant Compute DL380a Gen12 servers with NVIDIA L40S GPUs for inference and evaluation workloads.[23] The deployment was built in collaboration with HPE, NVIDIA, and STN (an HPE partner) through HPE's GPU One service.[23] CoreSite's CH2 data center in Denver provides co-location infrastructure supporting Skild's AI factory, which integrates both training and production inference capabilities.[39]
For simulation, Skild AI is a heavy NVIDIA shop. The company runs large-scale physics simulation on NVIDIA Isaac Lab and NVIDIA Omniverse, uses NVIDIA Cosmos for synthetic data augmentation, and trains on NVIDIA accelerated computing.[22] The case study NVIDIA published with Skild AI describes simulations with thousands of concurrent robot instances and references the ability to acquire "a millennium of experience within days."[22] That tooling overlap, plus NVentures' participation in the Series B and Series C, helps explain why NVIDIA is a strategic partner rather than a direct competitor, even though NVIDIA itself sells Isaac GR00T as a foundation model platform for humanoids.
At the edge, Skild deployments run on NVIDIA Jetson hardware embedded in partner robots.[7] This split between cloud training and on-device inference is similar to the pattern used by self-driving stacks and by other embodied AI platforms, but is unusual in that a single trained brain runs across so many different physical bodies in production.
The central product is the Skild Brain, marketed as the first unified robot foundation model.[37] It follows a hierarchical two-tier architecture: a low-frequency high-level policy reasons about the task and scene and provides inputs to a high-frequency low-level policy that converts those directives into precise joint angles and motor torques.[9] The Skild Brain is exposed to enterprise customers through an API that abstracts low-level control and lets developers build applications on top of robots without managing kinematics or sim-to-real themselves.
Three application platforms are layered on the same brain: a security and inspection platform with quadruped robots that navigate unstructured industrial environments, a mobile manipulation platform with bases and arms that can grasp, hand objects to people, and navigate, and an autonomous packing application targeted at warehouses.[9]
In late July 2025, Skild AI publicly unveiled the Skild Brain for the first time with an extended demonstration video.[9][37] Before this unveiling the company had operated largely in stealth on the product side despite the high-profile Series A funding.[24] The demonstration footage showed quadrupeds climbing stairs under adversarial conditions, fine-grained manipulation tasks including assembling small electronics components and AirPods into their charging case, bagel preparation and dishwashing, outdoor locomotion over uneven terrain, and a quadruped sitting on a concrete ledge.[9] CEO Pathak said: "Skild AI models can not only solve these easy tasks but also solve everyday hard tasks such as climbing stairs even under adversarial conditions, or assembling fine-grained items."[9]
The unveiling also described in-context adaptation behavior: when the Skild Brain is introduced to a new robot body or an unfamiliar environment where its actions fail, it adjusts behavior based on live in-context experience rather than requiring retraining.[9] One demonstration from the engineering team showed that when the brain was turned on while a quadruped was already standing upright, the model elected to operate the robot as if it were a small humanoid.[9] This was presented as evidence of in-context body identification rather than canned per-robot logic.
The company reported task performance in the 60% to 80% range within hours of data collection on new tasks, and recovery from mechanical failures such as a jammed wheel within a few seconds and a broken leg after a handful of attempts.[9]
Skild has publicly described several safety constructs baked into the Skild Brain. The model includes power and torque limits enforced at the low-level controller, intended to prevent any robot under its control from exerting force beyond a configured envelope even when the high-level plan would otherwise produce a forceful motion. Vision-based human awareness is used to widen safety margins when a person is detected in the workspace. The company has acknowledged that safety guarantees in fully unstructured environments such as homes remain an open research problem and has been explicit that its near-term deployments are confined to industrial and semi-structured commercial settings precisely because those environments allow physical layouts and procedural barriers to handle the long tail of failure modes. This is one of several ways the company's roadmap differs from full-stack humanoid competitors that have publicly demonstrated home deployments.
In March 2026 Skild AI announced a set of partnerships with major industrial and robotics companies that represented its first publicly named large-scale commercial deployments.[7][20]
The most concrete announcement involved Foxconn's manufacturing facility in Houston, Texas. Skild AI will deploy its omni-bodied brain to control dual-arm robots on Foxconn's NVIDIA Blackwell GPU server production lines, performing complex assembly operations including picking, placing, and drilling, tasks that currently require human operators.[25] The Foxconn deployment involves long-horizon task sequences spanning several minutes with on-the-fly recovery adjustments.[25] CEO Pathak framed the deployment as follows: "By training an omni-bodied intelligence that transfers skills across embodiments and environments, we're shifting from programming tasks to building systems that continuously learn and improve, even during deployment."[7] Pathak has also said publicly that partnering with industrial OEMs that operate hundreds of thousands of robots is how Skild expects to bootstrap its data flywheel at scale, rather than building a single proprietary humanoid platform.[7]
Simultaneously, Skild AI announced integration partnerships with ABB Robotics and Universal Robots (a Teradyne Robotics subsidiary).[7][20] ABB Robotics president Marc Segura stated the integration would help customers "scale industrial-grade automation more quickly" without task-by-task reprogramming.[20] Universal Robots CEO Jean-Pierre Hathout said the combination would bring "advanced AI capabilities to cobots" for handling variable tasks across industries.[20] Through Universal Robots, the partnership also extends to Mobile Industrial Robots (MiR), another Teradyne subsidiary that makes autonomous mobile robots for factory logistics.[20]
All three integration partners will embed the Skild Brain as a shared intelligence layer deployable across their robot portfolios. The NVIDIA toolchain, including Isaac Lab, Isaac Sim, Newton physics engine, and Cosmos world foundation models, underpins the simulation-to-deployment pipeline across all three relationships.[7] NVIDIA's Jetson hardware handles on-device inference at the edge.[7]
The brain runs across a broad set of robot bodies in Skild's public demos and deployments. The table below summarises the platforms that have been publicly shown or named by the company and its partners.
| Platform class | Example bodies | Use cases | Status |
|---|---|---|---|
| Quadrupeds | Unitree B2 and similar legged platforms | Security patrol, industrial inspection, stair climbing | Production deployments at named partners |
| Humanoids | Partner humanoids (not publicly named) | Mobile manipulation, light assembly | Demonstrated; commercial pilots underway |
| Industrial arms | ABB IRB series, partner cobots | Pick-and-place, assembly, drilling on Foxconn Blackwell lines | Partner integration announced March 2026 |
| Collaborative arms | Universal Robots UR series | Variable cobot tasks, light industrial automation | Partner integration announced March 2026 |
| Autonomous mobile robots | Mobile Industrial Robots (MiR), legacy Fetch fleet from Zebra | Material movement, warehouse logistics | Deployed via MiR partnership and Zebra acquisition |
| Mobile manipulators | Custom dual-arm bases | Hand-offs, picking, packing in warehouses | Internal platform; pilots with enterprise customers |
This breadth across morphologies is the strongest single piece of empirical support for the omni-bodied thesis. None of Skild's full-stack humanoid competitors have publicly demonstrated their model running on this many distinct body types, which is part of why Skild's pitch to integration partners has gained traction quickly.
As of late 2024 the company had no publicly named customers. By January 2026 Skild AI reported approximately $30 million of revenue accrued during 2025, with deployments described across security, construction, last-mile delivery, robotics in data centres, warehouses, and factory assembly.[2][10] The Series C investor list includes a number of strategic corporates that double as likely deployment partners, notably Samsung, LG, Schneider Electric, Salesforce Ventures, and CommonSpirit (a US healthcare system).[2]
The Foxconn assembly line deployment announced in March 2026 is the first publicly named mass deployment after several years of enterprise testing.[25] Prior deployments in warehousing, construction, and inspection were confirmed by the company but customer names were not disclosed.
In April 2026 Skild AI acquired Zebra Technologies' Robotics Automation business, including the Symmetry Fulfillment orchestration platform that coordinates fleets of robots in warehouses.[6][17][32] The financial terms were not disclosed, though Zebra received an equity stake in Skild AI as part of the deal.[11][17]
The acquired business traces back to Fetch Robotics, an autonomous mobile robot company founded in 2014 by Melonee Wise that was one of the early commercial pioneers in warehouse AMR logistics.[11][38] Zebra Technologies acquired Fetch in July 2021 for $291 million.[38] In December 2025, facing mounting losses, Zebra announced it was winding down the Fetch robotics division, with expected one-time pre-tax charges of up to $80 million (including roughly $60 million in non-cash asset impairment).[38] Skild's acquisition of the business before full wind-down preserved the team and technology.[11]
CEO Pathak cited the Fetch team's deployment experience as the primary acquisition rationale: "the Fetch Team is the main reason for the acquisition as they bring years of deployment experience."[11] The acquisition combines the omni-bodied AI brain with a deployed warehouse fleet management product, allowing customers to plug Skild Brain-driven humanoids, quadrupeds, autonomous mobile robots, and arms into existing fulfillment workflows through the Symmetry orchestration layer.[6] Skild plans to continue supporting existing Fetch/Zebra customers, sell new Fetch robots, and build advanced manipulation solutions that embed Skild Brain capabilities.[11]
The combination creates what Skild describes as the first end-to-end warehouse automation offering: humanoids for pick-and-place, quadrupeds for inspection, robotic arms for packing, AMRs for material movement, and a single orchestration layer controlling them all.[6]
| Company | Founded | Leadership | Stack | Disclosed funding | Notes |
|---|---|---|---|---|---|
| Skild AI | 2023 | Deepak Pathak (CEO), Abhinav Gupta (President) | Software-only foundation model (Skild Brain) | $1.835B+ | Omni-bodied, hardware-agnostic |
| Physical Intelligence | 2024 | Karol Hausman, Sergey Levine, Brian Ichter | Software-only foundation model (pi0, pi-0.5) | $1.1B+ | Most direct software-only peer |
| Covariant | 2017 | Pieter Abbeel | Manipulation foundation model RFM-1 | ~$625M; Amazon licensed in 2024 | Effectively absorbed into Amazon |
| Sanctuary AI | 2018 | Geordie Rose | Humanoid Phoenix + AI controller Carbon | $200M+ | Hardware plus model |
| Intrinsic | 2021 | Wendy Tan White | Industrial robot programming inside Alphabet | Internal | Vertical industrial automation |
| 1X Technologies | 2014 | Bernt Bornich | Humanoids (NEO, EVE) + model | ~$200M | Hardware plus model |
| Figure AI | 2022 | Brett Adcock | Humanoid robots + model (Helix) | Multi-billion | Hardware plus model |
| Tesla Optimus | 2021 | Elon Musk | Humanoid robot + in-house model | Internal | Hardware plus model |
| NVIDIA Isaac GR00T | 2024 | Jensen Huang and team | Foundation model platform + simulation tooling | Internal | Sells tooling to roboticists |
| Google DeepMind RT-2 / Gemini Robotics | 2023 | Demis Hassabis and team | Vision-language-action models | Internal | Research and partner integration |
| ABB Robotics | Legacy | Marc Segura | Industrial robots + Skild Brain integration | Internal (public company) | OEM partner; embedding Skild Brain |
| Universal Robots | Legacy | Jean-Pierre Hathout | Cobots + Skild Brain integration | Internal (Teradyne subsidiary) | OEM partner; embedding Skild Brain |
The closest comparable to Skild AI in positioning is Physical Intelligence, also a software-only robotics foundation model company spun out of Berkeley and Google DeepMind alumni. Both companies are training a single generalist model intended to drive many bodies; Physical Intelligence has open-sourced its pi0 model, while Skild AI has kept the Skild Brain proprietary. Both have Bezos Expeditions and Sequoia in their cap tables. Covariant is a useful reference for what happens when a robot foundation model company tries to commercialise: in August 2024 Amazon hired the founders and licensed the technology, effectively ending Covariant as a stand-alone player. Sanctuary AI, 1X, Figure, and Tesla are full-stack, with the model being one component in a larger humanoid robot product. Intrinsic at Alphabet attacks a narrower problem (programming industrial robots), while NVIDIA and Google DeepMind sell tooling and models to others rather than running a robotics product company directly.
The March 2026 announcement that ABB Robotics and Universal Robots are embedding the Skild Brain into their products is a meaningful competitive shift: it converts two large incumbent OEMs from potential competitors (as they might develop their own AI) into distribution partners.[7][20]
The three software-only robotics foundation model efforts that get compared most often are Skild's Skild Brain, Physical Intelligence's pi family, and Covariant's RFM-1. They differ along several axes.
| Axis | Skild Brain | Physical Intelligence pi-0.5 | Covariant RFM-1 |
|---|---|---|---|
| Body coverage | Quadrupeds, humanoids, industrial arms, cobots, AMRs | Mostly bimanual mobile manipulators; some humanoids | Industrial pick arms |
| Primary data source | Large-scale procedural simulation plus internet video plus teleop | Teleoperation on a small set of platforms, augmented with web video | Pick-and-place data from Amazon and partner warehouses |
| Release model | Closed, API access for enterprise customers | Open weights for research with commercial license terms for pi0 | Effectively internal to Amazon after 2024 licensing deal |
| Go-to-market | OEM partnerships, direct enterprise, warehouse and factory deployments | Direct enterprise pilots; selective hardware partnerships | Distribution through Amazon and partner warehouses |
| Investor base | SoftBank, NVIDIA, Lightspeed, Coatue, Bezos, Sequoia | Bezos, Sequoia, Lux, Thrive, OpenAI Startup Fund | Index, Radical, Amazon (later licensee) |
The most consequential distinction is on data. Physical Intelligence has bet on teleoperation as the primary lever, arguing that hours of high-quality human demonstrations on a manageable number of platforms produce a more reliable model than enormous quantities of simulated data. Skild has bet on simulation at extreme scale (the 100,000 morphologies and "a millennium of experience within days" framing) with teleop and video as supplementary signals.[22] Both teams are well-credentialed and the empirical question of which mixture wins is unresolved. Covariant's RFM-1 was trained on the narrowest data slice of the three (warehouse picking) and reached commercial deployment earliest, which is part of why it became attractive enough for Amazon to license the technology and absorb the founders rather than acquire the company outright.
Skild AI's pitch carries two big strategic bets. The first is that hardware will commoditise faster than people expect. If humanoid OEMs (Figure, Apptronik, Unitree Robotics, 1X, UBTECH, and others) actually ship at scale, the operator of the brain will be in a similar position to a cloud or operating-system layer above competing devices, the same logic Microsoft and Google used in PCs and phones. If hardware turns out to be the harder problem and one or two integrated players win in the same way Apple did, a software-only brain is in a weaker bargaining position. The second bet is that a single omni-bodied model can really beat platform-specific systems, on the theory that the same mathematics that let large language models exhibit emergent capabilities once trained on enough text should apply to a model trained on enough physical experience.
Skild's go-to-market is enterprise-led: data centre operators, warehouses, manufacturers, healthcare systems, and security customers who need physical labour with intelligence and either do not want to vertically integrate their own robotics stack or are happy to mix robot brands as long as one brain runs them. The Zebra Technologies acquisition is the clearest signal of that direction: it converted Skild from a pure software company into a software company with a deployed fleet management platform and a customer book in fulfillment.[6]
The partnerships with ABB Robotics and Universal Robots announced in March 2026 extend the distribution reach further.[7] ABB Robotics operates a global installed base of hundreds of thousands of industrial robots across automotive, electronics, and food and beverage manufacturing. Universal Robots is the world's largest cobot maker by units shipped. Embedding Skild Brain into both portfolios gives the model access to a far larger deployment surface than Skild could build organically.[20]
At Davos in January 2026, Pathak argued that the most economically consequential effect of generalised robot brains will not be in heavy industry but in the routine cognitive work that has historically been done at desks.[31] His position is that physical AI extends the reach of automation into categories of work that language-only AI cannot touch (loading, sorting, picking, inspection, maintenance) and that the combination of generative AI for office tasks and physical AI for the rest of the economy will compress timelines for labour disruption far more than either category would alone.[31] The remark drew significant press attention, partly because it inverted the conventional framing that white-collar work is less automatable than manual work.[31] Skild's commercial materials usually cite figures of roughly 2.1 million unfilled US manufacturing jobs by 2030, and Pathak has framed those gaps as the immediate addressable market while consumer and home applications remain on the longer-term roadmap.
Skild AI is part of a 2023 to 2026 wave of robotics startups built around the idea that the recipe that produced large language models will work for robots. The wave was set up by scaling in deep reinforcement learning, the appearance of vision-language-action models like RT-2 and PaLM-E, NVIDIA pushing simulation and a Isaac GR00T platform for humanoids, falling humanoid hardware costs, and renewed VC enthusiasm for embodied AI after the success of generative AI in software. Investors that had concentrated on cloud and consumer software started writing big robotics cheques: Lightspeed, Coatue, SoftBank Vision Fund, Bezos Expeditions, Sequoia, Founders Fund, Khosla, Thrive, Lux, and Felicis are all visibly active in this space. The macro pitch shared by most of these companies is chronic labour shortages in manufacturing, warehousing, healthcare, and construction; the Series A press materials cited around 2.1 million unfilled US manufacturing jobs by 2030.[1]
There are real reasons to be cautious. Generalisation across robot bodies is genuinely hard: demos in which a model adapts after a robot loses a leg are striking, but production deployment needs robust performance over many months under conditions that look nothing like a demo, and moving from 60% to 80% task success in a hand-picked demo to 99% reliability in front of a paying customer is the work of years. Real-world safety and trust matter more for embodied AI than for chat; a chatbot hallucination is annoying, but a robot that grabs the wrong object near a person can hurt someone. Regulatory regimes around safety-critical embodied AI are still in their infancy.
Competitive pressure is intense. Physical Intelligence is open-sourcing aggressively, NVIDIA owns the simulation infrastructure layer and ships its own foundation model platform with Isaac GR00T, Google DeepMind keeps publishing strong vision-language-action results, hardware-first humanoid companies have their own AI teams, and incumbents in industrial automation (ABB, Fanuc, KUKA, Honda) are not going to stand still.
Valuation also outruns revenue: at a $14 billion valuation against around $30 million of 2025 revenue, the company is priced almost entirely on its expected long-run market share of robot intelligence, not on near-term financials.[2][15] That is normal for the category but is a real risk if scaling laws for robotics turn out to be slower or messier than they did for text. There is also a substantive question about whether truly omni-bodied models actually beat well-funded specialist models on commercially important tasks; critics inside the field argue that a model tuned for a single embodiment will always have a quality edge for that embodiment.
The SEC Form D filing discrepancy noted by Technical.ly is worth flagging: as of the July 2025 Skild Brain unveiling, Form D filings under Skild-related entities showed only about $1 million under a special purpose vehicle, while the company was publicly claiming to have raised $814 million (per PitchBook).[24] The explanation is likely structural (tranches filed under different entity names or structured investment vehicles), but the opacity is unusual for a company of this scale.
Beyond the commercial risks, there are open scientific questions that Skild's public materials do not yet resolve. The company has not published a peer-reviewed paper describing the Skild Brain architecture, the training mixture composition, or the evaluation methodology in detail, which makes it hard for outside researchers to assess claims like "60% to 80% task success within hours of data collection" or "a millennium of experience within days." Comparable robotics models from Google DeepMind and Physical Intelligence have published more technical detail. Critics in the academic robotics community have noted that demos showing in-context body identification can be partly explained by simple proprioceptive feedback loops that do not require a large foundation model, and have asked for benchmark results on standardised cross-embodiment evaluation suites. Skild has not yet published such results.
A second open question concerns the limits of simulation. The Isaac Lab and Omniverse stacks have improved rapidly but still suffer from the well-known sim-to-real gap in contact-rich manipulation, where small differences in friction, compliance, and sensor noise produce large divergences between simulator predictions and physical outcomes. Skild's response is to mix simulation with real-world teleop data and to rely on in-context adaptation in deployment, but the question of how much real-world data is required to close the gap for a given task is unresolved and is one of the central scientific bets the company is making.
| Date | Event |
|---|---|
| May 2023 | Company founded by Deepak Pathak and Abhinav Gupta |
| July 9, 2024 | Series A of $300 million announced at $1.5 billion valuation; emergence from stealth with first public demo videos |
| March 2025 | HPE partnership announced; Skild AI deploys private AI-as-a-service compute on HPE Cray XD670 with NVIDIA H200 GPUs |
| April 2025 | Series B of approximately $135 million (SEC-reported) at $4.5 billion to $4.7 billion valuation, led by SoftBank with NVIDIA, Samsung, LG, Bezos, Schneider Electric, CommonSpirit, and others |
| July 30, 2025 | Public unveiling of the Skild Brain with extended demonstrations of stair climbing, fine-grained manipulation, and mechanical failure recovery |
| January 14, 2026 | Series C of $1.4 billion at more than $14 billion post-money valuation, led by SoftBank with NVentures; revenue disclosed at roughly $30 million for 2025 |
| January 21, 2026 | Pathak speaks at the World Economic Forum in Davos, framing physical AI as the foundation of AGI and predicting desk-job disruption |
| February 2026 | Bengaluru office opened, Skild's first international office |
| March 16-19, 2026 | Partnerships with NVIDIA, Foxconn, ABB Robotics, and Universal Robots announced; Foxconn Blackwell GPU assembly line deployment confirmed |
| April 15, 2026 | Acquisition of Zebra Technologies' Robotics Automation business (formerly Fetch Robotics) and Symmetry Fulfillment orchestration platform |
The research career is what makes Skild AI's pitch credible. Pathak's curiosity-driven exploration paper has been cited tens of thousands of times and is one of the foundational pieces of intrinsic motivation literature.[35] His work on self-assembling morphologies set up the cross-embodiment framing that the Skild Brain is now scaling, and his later applied papers on legged locomotion (in particular rapid motor adaptation for quadrupeds) helped make the case that learning-based controllers could ride out the real-world distribution shifts that defeated earlier model-based approaches.[36] Gupta's work on self-supervised visual learning, large-scale grasping datasets, and visual common sense did similar conceptual heavy lifting on the perception side, and his leadership of the FAIR Pittsburgh robotics lab from 2018 to 2022 helped establish Pittsburgh as a centre of robot-learning research outside the Bay Area.[34] Skild AI is essentially saying that the methods that worked in their labs at academic scale will keep working at industrial scale with enough simulation, real-world data, and compute. Whether that turns out to be right is still an open question, but the people making the bet are demonstrably the people who developed many of the methods the rest of the field is also using.