Skild AI Deep Dive: A General Robot Foundation Model Incubated at CMU's Robotics Institute, With a USD 1.4B Series C at a USD 14B Valuation, a Conflicting Series B Size, and an ARR Figure That Is a Run-Rate Rather Than Revenue

Executive Summary

Skild AI, Inc. is a Pittsburgh-based company that builds a general-purpose robot foundation model, incubated at Carnegie Mellon University's Robotics Institute and founded in 2023. It is one of the best funded and most technically credible companies in the wave of general-purpose robot models, and it is also one of the companies where the gap between headline claims and auditable evidence is widest. This report treats those two facts as equally important rather than as a reason to prefer either.

The scientific leadership is the strongest asset and it is verifiable. Deepak Pathak, chief executive, is a Raj Reddy Associate Professor at Carnegie Mellon, holds a doctorate from the University of California Berkeley, authored the 2017 curiosity-driven exploration work that has been cited more than four thousand times, was named an MIT Innovator Under 35 in 2024, and received a Sloan Research Fellowship in 2025. Abhinav Gupta, president, leads the Carnegie Mellon Robotics Institute, was a founding member of Meta's Fundamental AI Research team, and holds a doctorate from the University of Maryland. The two previously worked together at Meta FAIR, and the company's argument that it is unusually well placed to build a general robot model rests on exactly this pairing.

The technical arc has three dated milestones and they are worth separating, because they make different promises. Skild Brain was first released on 29 July 2025 and was positioned as a hierarchical controller spanning many robot embodiments; on 24 September 2025 the company described training it to control a universe of one hundred thousand robot types. On 19 March 2026 the company announced collaborations with ABB Robotics, Universal Robots and NVIDIA. On 15 April 2026 it acquired the robotics business of Zebra Technologies, which includes Fetch Robotics and its Symmetry Fulfillment line. The S1 model arrived on 18 August 2026 with a specific in-context learning claim: a single video demonstration is sufficient to execute an unseen task for up to roughly ten minutes, with model weights frozen during that execution, trained on simulation data, internet human video, gripper footage and a small amount of teleoperation. The most concrete internal numbers are a step success rate of about 66% on unseen tasks against a 9% baseline for a language-conditioned approach, and about 96% on seen tasks. On 23 September 2026 the company introduced Physical Self-Play, in which the only objective is scoring a goal, and the policy self-plays inside NVIDIA Isaac Sim for the equivalent of 140 simulated years before transferring to a real humanoid.

The commercial picture is the strongest quantitative evidence in the file, and it needs careful reading. On 10 September 2026 the company announced that annual recurring revenue had passed USD 100 million, ten months after the first commercial deployment, with more than sixty paying customers and USD 50 million of recognised revenue. The composition matters: mobility accounts for roughly ten percent of revenue, of which autonomous mobile robots and the acquired Zebra business contribute four percent, and the remaining ninety percent comes from manipulation. The critical caveat is that a USD 100 million ARR against USD 50 million of recognised revenue is a forward run-rate figure, not booked business, and the gap between the two is the single most important number to hold in mind when valuing this company.

The financing is enormous and the documentation has a hole in it. The Series A on 9 July 2024 raised USD 300 million at a USD 1.5 billion valuation, led by Lightspeed, Coatue, SoftBank and Bezos Expeditions, with Sequoia's Stephanie Zhan joining the board. The Series B in June 2025 was led by SoftBank, with NVIDIA investing USD 25 million, Samsung USD 10 million, and Korean investors LG, Hanwha and Mirae Asset each investing between USD 5 million and USD 10 million, at a reported valuation of about USD 4.5 billion. The size of that round is reported two ways that differ by nearly four times, USD 135 million against USD 500 million, and this report does not resolve the difference. The Series C on 14 January 2026 raised USD 1.4 billion at a valuation above USD 14 billion, led by SoftBank with NVentures, Macquarie and Bezos Expeditions participating and existing investors doubling down, with LG, Schneider Electric, CommonSpirit Health, and Salesforce Ventures joining as new investors. Cumulative funding is stated as more than USD 1.8 billion to 1.83 billion by official and database sources, as more than USD 2 billion by the chief executive, and totals roughly USD 2.2 billion if the individual rounds are added directly.

The evidence base for reliability is thin, and the company itself appears to know it. No failure rate, task success rate on customer deployments, or failure mode taxonomy has been published. The 66% unseen-task step success rate is an internal test figure. The Physical Self-Play announcement did not publish a real-world goal-scoring rate, which is the number that would actually matter for a deployment decision.

The competitive set is the strongest cohort in robotics. Physical Intelligence, Google DeepMind with Gemini Robotics, NVIDIA with GR00T, Figure AI and 1X are all competing for the same budgets and, in several cases, the same customers. Skild AI's differentiation is the combination of a genuinely general model, a foundation-model training approach that is not tied to a single robot body, and a commercial track record that arrived earlier than most of its peers.

Company Overview

Skild AI, Inc. was founded in 2023 and is headquartered in Pittsburgh, with additional presence in San Francisco and Bangalore, the latter added in February 2026. The company was incubated at Carnegie Mellon University's Robotics Institute, and the two founders are the institute's academic leadership: Deepak Pathak as chief executive and Abhinav Gupta as president.

The founders' record is the clearest part of the company's claim. Pathak's 2017 work on curiosity-driven exploration established that an agent's own sense of novelty could substitute for hand-designed extrinsic rewards, and it has been cited more than four thousand times. He was an MIT Innovator Under 35 in 2024 and a Sloan Research Fellow in 2025. Gupta led the robotics laboratory at Facebook AI Research and was a founding member of FAIR, and his group's work on large-scale self-supervised learning for visual representation predates the current wave of embodied foundation models. The technical argument that a general robot model requires both exploration research and large-scale representation learning is, in this case, being made by the people who did that work.

The corporate structure has been made simpler by an acquisition. On 15 April 2026 the company acquired the robotics business of Zebra Technologies, which brought Fetch Robotics and the Symmetry Fulfillment platform into the group. This is a strategically coherent move in one direction, since Fetch was one of the more credible mobile manipulation stacks in the market, and a strategically ambiguous one in another, because it is a business line rather than a model capability.

Commercial disclosure is better than for most private robotics companies but remains incomplete. More than sixty paying customers, USD 100 million of annual recurring revenue on a run-rate basis, USD 50 million of recognised revenue, and a revenue composition of roughly ten percent mobility, four percent autonomous mobile robots and the acquired Zebra business, and ninety percent manipulation. There is no disclosed gross margin, no customer concentration, no retention or expansion data, and no breakdown of ARR between the newest and oldest accounts.

There is no public information on headcount, cash balance, burn rate, employee equity, or a listing timetable.

Common Misreadings: Five Errors That Change the Analysis

The USD 100 million ARR is not revenue. The company announced that annual recurring revenue passed USD 100 million on 10 September 2026, and separately reported USD 50 million of recognised revenue. These are two different quantities and the gap between them is the number that matters. Treating the ARR as booked revenue overstates the current business by a factor of two.

Skild AI did not train a humanoid from scratch. It acquired a robotics business from Zebra Technologies in April 2026, which included Fetch Robotics and Symmetry Fulfillment. A company that acquires rather than builds is not the same risk profile as one that has engineered everything internally, and the acquisition also brought four percent of revenue and a hardware-adjacent business that the model-first narrative does not cover.

The 66% figure is an internal test result on step success, not a deployment success rate. It comes from the company's own evaluation on unseen tasks, and it is a step-level metric. No customer deployment success rate has been published, and the two are not comparable.

The 140 simulated years are a training procedure, not an operating record. Physical Self-Play ran a policy against itself inside NVIDIA Isaac Sim for the equivalent of 140 simulated years. No real-world goal-scoring rate was published alongside it, so the transfer result remains qualitatively described.

The Series B was not a single well-defined event. The valuation is reported consistently at about USD 4.5 billion, but the amount raised is reported as either USD 135 million or USD 500 million. The valuation is the more reliable of the two figures because it appears in the official announcements.

Technology

The technical claim is that a single model can control a wide range of robot embodiments, learn new tasks from very little demonstration, and improve its own behaviour through self-play against a physical objective.

Skild Brain, released 29 July 2025, is the hierarchical controller layer. The design point is embodiment breadth: on 24 September 2025 the company described training the model to control a universe of one hundred thousand robot types. The significant part is the architecture rather than the count, since a hundred thousand embodiments is a statement about the interface abstraction over which the model operates, and the published record does not describe how many of those embodiments were evaluated on real hardware rather than in simulation. The March 2026 partnerships with ABB Robotics, Universal Robots and NVIDIA are what turn this breadth into a commercial channel: ABB and Universal Robots are two of the largest industrial robot manufacturers in the world, and NVIDIA is the dominant compute supplier for the field.

S1, released 18 August 2026, is the in-context learning contribution and it is the most technically interesting thing the company has published. The mechanism is that the model's weights are frozen and the new task specification is supplied in context, which for a robot means a video demonstration rather than a new training run. The training data is a mixture of simulation, human video from the internet, gripper footage, and a small amount of teleoperation. The published internal numbers are approximately 66% step success on unseen tasks against 9% for a language-conditioned baseline, and approximately 96% on seen tasks.

The 66% number deserves careful reading rather than celebration. It is a step-level success rate on unseen tasks, and its baseline of 9% is a deliberately weak comparison class, a language-conditioned baseline rather than the best available alternative method. The meaningful reference point is what a well-engineered scripted behaviour or a human teleoperator achieves on the same task distribution, and that comparison has not been published. A frozen-weight in-context system that succeeds roughly two thirds of the time at the step level on unseen tasks is a meaningful research result; it is not yet a deployment reliability claim.

Physical Self-Play, introduced 23 September 2026, is the most interesting methodological idea and the least quantified. The setup is deliberately minimal: the only objective is scoring a goal, and the policy trains itself by opposing itself inside NVIDIA Isaac Sim, in a process the company describes as the equivalent of 140 simulated years, before transfer to a real humanoid. The appeal is that it removes the need to specify a task distribution and reward function by hand, which is the main scaling bottleneck in robot learning. The announced figures do not include a real-world goal-scoring rate, and the policy was not trained to generalise across many tasks, only to score one.

The physics connection is real but indirect. The simulation in Physical Self-Play runs on NVIDIA Isaac Sim, which is built on the PhysX and Omniverse physics stack. A company whose self-play procedure depends on a simulator supplied by a third party inherits that simulator's fidelity limits, and that dependency is structural rather than cosmetic. More broadly, Skild AI's contribution is in learning and control rather than in physics modelling, which places it on the model side of the boundary that separates physical AI companies, and it is not a substitute for a solver, a mesh generator, or a multiphysics code.

Product and Commercial Model

The commercial model is a foundation model sold as a platform, with the deployment surfaces being industrial robots, mobile manipulators, and humanoids.

The disclosed revenue picture is the strongest available. More than sixty paying customers, USD 100 million of annual recurring revenue on a run-rate basis, and USD 50 million of recognised revenue. The company reported the ARR milestone on 10 September 2026, ten months after the first commercial deployment, which is a genuinely fast ramp on any reading.

The composition matters more than the total. Mobility is roughly ten percent of revenue, of which autonomous mobile robots and the acquired Zebra business account for four percent, and roughly ninety percent is manipulation. This means the business is, for now, a manipulation company, and the humanoid and physical world model narrative is an option on future mix rather than the present mix.

The named customer references are substantive and physical. NVIDIA and Foxconn on dual-arm assembly for Blackwell systems, Sumitomo Wiring Systems on wire harness work, Mitsui on a commercial kitchen pilot in Japan serving 1.4 million meals a day, LG CNS, and the Pittsburgh International Airport operating a quadruped. These are industrial deployments with production consequences, which is a stronger form of evidence than a demonstration video.

No pricing is published, no gross margin is disclosed, and there is no renewal or expansion data. With sixty-plus paying customers, the concentration question is material and cannot be answered from public sources.

The acquisition of Zebra's robotics business changes the product story in a way that is not yet reflected in the narrative. Fetch Robotics and Symmetry Fulfillment bring a working commercial product with existing customers, which is a reasonable way to buy revenue and channel, and it also introduces a hardware-adjacent business and a four percent revenue contribution that sits outside the model thesis.

Competitive Landscape

Skild AI competes in the most heavily funded technical field in robotics, and its competitive position is stronger than the average company in that field but weaker than the headline valuation implies.

Against Physical Intelligence, the comparison is the closest one. Both pursue general-purpose robot models with a foundation model framing, both emerged from academic research, and both have raised at the billion-dollar scale. Physical Intelligence has the earlier commercial visibility in research and developer channels; Skild AI has the earlier revenue.

Against Google DeepMind with Gemini Robotics, the comparison is one of resources against focus. Google DeepMind can fund a general robot model indefinitely and has both the robotics hardware knowledge and the deployment channel; Skild AI's advantage is that it is not balancing a robot business against a search business.

Against NVIDIA with GR00T, the relationship is more entangled than competitive. NVIDIA is an investor in Skild AI and the supplier of both the compute and the Isaac Sim simulator on which the self-play procedure depends, and it announced a collaboration with Skild AI in March 2026. A competitor that is also a supplier and a shareholder is a genuine strategic ambiguity, and it is a pattern across the field rather than unique to this company.

Against Figure AI and 1X, the comparison is that those two built hardware and a humanoid, while Skild AI built a model and bought a robotics business. If hardware is where the durable value ends up, Skild AI's strategy is exposed; if the model is, its position is strong. The published evidence does not settle this question, and the debate is itself the investment risk.

Against industrial robot incumbent platforms, the ABB and Universal Robots relationships are the strategic core of the go-to-market. Embedding in the dominant industrial robot installed base is a distribution advantage that no startup without those partnerships has.

Financing

The Series A on 9 July 2024 raised USD 300 million at a USD 1.5 billion valuation. It was led by Lightspeed Venture Partners, Coatue Management, SoftBank and Bezos Expeditions, with Sequoia Capital's Stephanie Zhan joining the board. The size and the valuation were reported consistently.

The Series B in June 2025 was led by SoftBank, with NVIDIA investing USD 25 million and Samsung USD 10 million, and LG, Hanwha and Mirae Asset each investing between USD 5 million and USD 10 million. The valuation is reported at approximately USD 4.5 billion.

The size of the Series B is reported two ways and the difference is not resolved here. One record gives USD 135 million for the round, which would make it the total including the strategic amounts. Another gives USD 500 million. The valuation figure appears in the official announcements and is the more reliable of the two; the amount is reported in both a database record and a press account. Both are carried below.

The Series C on 14 January 2026 raised USD 1.4 billion at a valuation above USD 14 billion. SoftBank led, with NVentures, Macquarie and Bezos Expeditions participating, existing investors doubling down, and LG, Schneider Electric, CommonSpirit Health and Salesforce Ventures entering as new investors.

Cumulative funding has three published versions. Official and database sources give more than USD 1.8 billion to 1.83 billion. The chief executive has said more than USD 2 billion. Summing the individual rounds as reported gives approximately USD 2.2 billion. All three are reported below and the difference between them is a matter of which rounds are counted and whether the amounts are gross or net.

The seed round is also documented two ways. One record gives USD 14.5 million from Lightspeed and Sequoia, and another gives no public amount.

DISCREPANCIES

1. Series B size. Convention A: USD 135 million (DB). Convention B: USD 500 million (MEDIA). Ruling: both are reported, and the valuation of approximately USD 4.5 billion is treated as the more reliable part of that round because it appears in official announcements. The two amounts differ by a factor of nearly four, and no source explains whether one includes a tranche the other excludes. No cumulative total is presented as settled.

2. Cumulative funding. Convention A: more than USD 1.8 billion to 1.83 billion (OFFICIAL and DB). Convention B: more than USD 2 billion (the chief executive's statement). Convention C: approximately USD 2.2 billion from summing the reported rounds. Ruling: all three are reported and none is presented as the answer. The spread is explained by which rounds are counted and whether announced amounts are gross proceeds or amounts actually received. This report does not select a single cumulative figure.

3. Seed round. Convention A: USD 14.5 million from Lightspeed and Sequoia (DB). Convention B: no public amount (OFFICIAL disclosure, amount not stated). Ruling: the amount is reported as a database figure and the absence of an official amount is stated explicitly. The existence of the round is not in dispute.

4. ARR against recognised revenue. Convention A: annual recurring revenue above USD 100 million as of 10 September 2026 (OFFICIAL). Convention B: recognised revenue of USD 50 million (OFFICIAL). Ruling: both are reported and the distinction is stated as the central reading instruction for this company. The ARR is a forward run-rate and the recognised revenue is booked business, and a reader who uses the larger number as current revenue will overstate the business by a factor of two.

5. Revenue composition. Convention A: mobility approximately ten percent of revenue (company disclosure). Convention B: autonomous mobile robots plus the acquired Zebra business approximately four percent, with the remaining approximately ninety percent from manipulation. Ruling: both are reported together and are read as a single consistent breakdown. The two statements are complementary rather than contradictory, and together they establish that the present business is predominantly manipulation.

6. Unseen-task performance. Convention A: approximately 66% step success on unseen tasks and approximately 96% on seen tasks (company internal evaluation). Convention B: a 9% baseline for a language-conditioned approach (company comparison). Ruling: all three figures are reported, with the metric definition stated as step-level success on an internal test. The baseline is identified as language-conditioned rather than as the best available alternative method, and no customer deployment success rate has been published. No third-party reproduction exists.

7. Physical Self-Play. Convention A: self-play inside NVIDIA Isaac Sim for the equivalent of 140 simulated years before transfer to a real humanoid (company announcement, 23 September 2026). Convention B: no real-world goal-scoring rate published alongside the announcement. Ruling: the procedure is reported and the absence of a real-world quantification is stated as part of the finding. This is a case of one disclosure rather than of two conflicting accounts.

8. Robotics acquisition. Convention A: the company acquired the robotics business of Zebra Technologies on 15 April 2026, including Fetch Robotics and Symmetry Fulfillment (company announcement). Convention B: no source indicates that the acquired hardware and products were developed internally by Skild AI. Ruling: the acquisition is reported as an acquisition, and the acquired assets are not counted as internally developed capability. The acquisition is treated as revenue and channel, not as technical validation of the model.

9. Robot embodiment breadth. Convention A: Skild Brain trained to control a universe of one hundred thousand robot types (company statement, 24 September 2025). Convention B: no disclosure of how many of those embodiments were evaluated on physical hardware rather than in simulation. Ruling: the figure is reported as a company statement with the disclosure gap stated. The count is treated as a statement about the interface abstraction rather than as a measured deployment breadth.

10. Headcount and financial discipline. Convention A: no public headcount, cash balance, burn rate, or listing timetable. Convention B: a valuation above USD 14 billion at the January 2026 Series C. Ruling: the absence of disclosure is reported as a finding in the risk section rather than as an absence of information. The valuation is a market observation about a private round and does not substitute for a disclosed balance sheet.

Risks and Due Diligence Gaps

The reliability evidence is the largest gap. No failure rate, no customer deployment success rate, and no failure mode taxonomy has been published. The 66% unseen-task step success rate is an internal test against a weak baseline class. For a company whose customers are wiring harness assembly and commercial kitchen work, the number that decides the purchase is the deployment success rate, and it is not public. Due diligence action: a customer reference call programme covering at least six deployments across at least three industries, and the internal failure taxonomy.

The ARR headline and the revenue reality diverge by a factor of two. USD 100 million of run-rate against USD 50 million of recognised revenue, ten months after the first commercial deployment, means the business either has an unusual concentration of recent deployments or that the run-rate includes commitments that have not converted. Either way, the run-rate is not revenue and the two should be reported as a pair.

The model-first thesis depends on the simulator being good enough, and the simulator is NVIDIA's. Isaac Sim, the PhysX and Omniverse stack, defines the fidelity ceiling on the self-play procedure. A model trained against an approximation of reality inherits that approximation's error, and Skild AI has no simulator of its own. The dependency is structural and also creates an awkward competitive position with a company that is simultaneously an investor, a supplier, and a rival.

The omni-bodied claim is a 2027-to-2028 proposition. The stated ambition is mass deployment of a model that transfers across embodiments, and the company's own framing places that at around 2027. No deployment-scale evidence exists for it, and the company has not published a roadmap against which progress could be tracked.

Revenue concentration in a sixty-customer base is unquantifiable. With more than sixty paying customers, one or two accounts could represent a large share of revenue, and the named references include NVIDIA and Foxconn on the same programme, which raises the question of whether these are separate contracts or a single programme counted twice. There is no concentration disclosure.

The valuation embeds a lot of optionality. A USD 14 billion valuation for a company with USD 50 million of recognised revenue implies a large multiple on a forward basis, which requires the humanoid and general-pretraining narratives to convert. If the general robot model market develops into a hardware-dominated market instead, this is the wrong company at this price.

Acquisition integration is unmonitored. The Zebra robotics business arrived in April 2026, roughly five months before this report, and contributes four percent of revenue. Integration of a hardware-adjacent business into a model company is a known source of distraction, and there is no public information on how it has gone.

The competitive field is capital-rich and moving fast. Physical Intelligence, Google DeepMind, NVIDIA, Figure and 1X are all well funded and all shipping. Being a credible participant in this field is not itself a moat, and the company's own customer list contains one of its principal competitors.

No exit route and no public market signal. The company is private with no listing timetable, no disclosed valuation methodology, and no investor liquidity event on record.

Overall Assessment

• Technology 8.0. A research record that can be verified at the level of individual citations, with a curiosity-driven exploration paper cited more than four thousand times and a MIT Innovator Under 35 award and a Sloan Fellowship. A general hierarchical controller spanning a stated hundred thousand embodiments, a genuine in-context learning contribution with frozen weights and single-video demonstration specifications, a self-play methodology that removes the need for hand-specified rewards, verified integration with NVIDIA Isaac Sim, and industrial partnerships with ABB Robotics and Universal Robots. Deducted points: the 66% unseen-task result is an internal step-level metric against a weak baseline class, no real-world quantification accompanies the self-play announcement, no third-party reproduction of any capability figure, and a simulator dependency on a competitor.
• Commercialisation 7.5. More than sixty paying customers, USD 100 million of run-rate revenue ten months after first deployment, USD 50 million of recognised revenue, a clear and business-relevant revenue mix of roughly ninety percent manipulation, and named production deployments with NVIDIA and Foxconn, Sumitomo Wiring Systems, Mitsui, LG CNS and the Pittsburgh International Airport. Deducted points: no pricing, no gross margin, no customer concentration, no retention or expansion data, and the ARR figure requires a run-rate caveat that is easy to lose.
• Financing 8.0. A USD 300 million Series A at USD 1.5 billion, a Series B at roughly USD 4.5 billion with SoftBank leading and NVIDIA, Samsung, LG, Hanwha and Mirae Asset all participating, and a USD 1.4 billion Series C at above USD 14 billion with SoftBank leading and NVentures, Macquarie, Bezos Expeditions, Schneider Electric and Salesforce Ventures participating. Cumulative funding above USD 1.8 billion on official and database figures. Deducted points: the Series B size conflict is unresolved, three published cumulative totals differ, the seed amount is not officially disclosed, and there is no disclosed valuation methodology or secondary liquidity event.
• Product maturity 7.5. A shipped general model with an in-context learning mode, an established hierarchical control layer, a partnership channel through the two largest industrial robot manufacturers, an acquired commercial robotics product line with existing customers, and a self-play training capability that no competitor has publicly described in the same terms. Deducted points: the flagship reliability number is internal, the omni-bodied claim has no deployment-scale evidence, the acquisition is too recent to have been assessed, and no pricing or product packaging is public.
• Market position 7.5. Positioned in the best-funded technical field in robotics with two of the field's most respected academic figures, an industrial channel through ABB and Universal Robots, a compute and simulator relationship with NVIDIA, and a commercial track record that arrived ahead of most peers. Deducted points: competing against Google DeepMind and NVIDIA with a fraction of the resources, an investor who is also a supplier and a competitor, hardware-strategy risk against Figure and 1X, no international or Chinese market presence evidence, and a valuation that prices a large amount of unproven optionality.

References

1. Skild AI company and product announcements, including the 29 July 2025 Skild Brain release and the 24 September 2025 statement on controlling one hundred thousand robot types
2. Skild AI S1 model announcement, 18 August 2026, including in-context learning from a single video demonstration for tasks of up to roughly ten minutes with frozen weights, the simulation and human video training mixture, and the approximately 66% and 96% internal success rates against a 9% language-conditioned baseline
3. Skild AI Physical Self-Play announcement, 23 September 2026, including the self-play procedure in NVIDIA Isaac Sim for the equivalent of 140 simulated years and the transfer to a real humanoid, with no real-world goal-scoring rate published
4. Skild AI announcements of 19 March 2026 regarding collaborations with ABB Robotics, Universal Robots and NVIDIA
5. Skild AI and Zebra Technologies acquisition announcement, 15 April 2026, covering the robotics business including Fetch Robotics and Symmetry Fulfillment
6. Skild AI ARR announcement, 10 September 2026, including the USD 100 million run-rate milestone, more than sixty paying customers, USD 50 million of recognised revenue, and the revenue composition of roughly ten percent mobility and four percent autonomous mobile robots and acquired Zebra business
7. Series A announcement, 9 July 2024, including USD 300 million at a USD 1.5 billion valuation led by Lightspeed, Coatue, SoftBank and Bezos Expeditions, with Sequoia's Stephanie Zhan joining the board
8. Series B records, June 2025, including the SoftBank lead, the NVIDIA, Samsung, LG, Hanwha and Mirae Asset participation, the approximately USD 4.5 billion valuation, and the USD 135 million against USD 500 million size conflict
9. Series C announcement, 14 January 2026, including USD 1.4 billion at a valuation above USD 14 billion, the SoftBank lead, participation by NVentures, Macquarie and Bezos Expeditions, and new investors LG, Schneider Electric, CommonSpirit Health and Salesforce Ventures
10. Database and press records for the seed round, including the USD 14.5 million from Lightspeed and Sequoia against the absence of a publicly disclosed amount, and for cumulative funding figures of more than USD 1.8 billion, more than USD 2 billion and approximately USD 2.2 billion
11. Skild AI deployment announcements naming NVIDIA and Foxconn on Blackwell dual-arm assembly, Sumitomo Wiring Systems on wire harness work, Mitsui on a Japanese commercial kitchen pilot serving 1.4 million meals daily, LG CNS, and the Pittsburgh International Airport quadruped
12. Carnegie Mellon University and University of California Berkeley records for Deepak Pathak, including the 2017 curiosity-driven exploration work, the Raj Reddy Associate Professorship, the MIT Innovator Under 35 distinction in 2024 and the 2025 Sloan Research Fellowship
13. Carnegie Mellon University and Meta FAIR records for Abhinav Gupta, including the Robotics Institute directorship, FAIR founding membership and the University of Maryland doctorate
14. NVIDIA Isaac Sim and PhysX documentation, used to establish the simulator fidelity dependency underlying the Physical Self-Play training procedure

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of October 23, 2026; please refer to official company announcements and filings.