Skild AI

Skild AI is a robotics company based in Pittsburgh, Pennsylvania, that develops general-purpose foundation models for robot control rather than robot hardware. It was founded in 2023 by Deepak Pathak and Abhinav Gupta, both professors at Carnegie Mellon University's Robotics Institute, around the position that robots will not deploy widely until a single model can drive many different bodies and tasks (Source: sequoiacap.com). Its stated aim, in Pathak's phrasing on his university page, is "any robot, any task, one brain" (Source: cs.cmu.edu). In January 2026 it raised a $1.4 billion Series C at a valuation above $14 billion (Source: techcrunch.com).

Snapshot

Valuation

Total capital raised exceeds $2 billion, a figure Pathak gave to Bloomberg at the time of the Series C (Source: techcrunch.com).

Founders and origin

Deepak Pathak is the Raj Reddy Associate Professor at Carnegie Mellon University, a member of the Robotics Institute and affiliated with the Machine Learning Department; he holds a PhD from UC Berkeley advised by Alexei Efros and Trevor Darrell, spent a year as a researcher at Meta AI Research, and is Skild's co-founder and chief executive (Source: cs.cmu.edu). Abhinav Gupta is the company's co-founder and president, a tenured professor at the Robotics Institute, and was a founding member and research leader of Meta's FAIR Robotics group in Pittsburgh (Source: sequoiacap.com). Both are graduates of IIT Kanpur, and Gupta holds a computer-science PhD from the University of Maryland (Source: cs.umd.edu).

The company emerged from stealth in July 2024 with the $300 million Series A. Sequoia, announcing its participation, described the two founders' complementary research lines — Pathak's work on curiosity-driven exploration and Gupta's on large-scale learning from unlabeled data — as the basis for the investment (Source: sequoiacap.com).

Business model

Skild does not sell robots. Its product is the Skild Brain, a foundation model licensed to robot manufacturers and enterprises as a control layer that is intended to work across humanoids, quadrupeds and robotic arms without per-platform engineering (Source: sequoiacap.com). The company's position, as reported at the Series C, is that the principal obstacle to robot adoption is the volume of task-specific training each new behavior requires, and that a model able to adapt during deployment removes it (Source: techcrunch.com).

Models

LocoFormer (September 2025)

LocoFormer is an in-context learner for locomotion that adapts by accumulating live experience in its prompt, described by the company as adapting "even when thrown into an unseen embodiment and environment." It was published as arXiv:2509.23745 by Min Liu, Deepak Pathak and Ananye Agarwal (Introducing S1: In-Context Learning for Robotics).

S1 (August 2026)

S1 is Skild's flagship manipulation model, released in a technical post dated August 18, 2026. Its design point is that a task is specified by a video demonstration placed in the model's context rather than by a language instruction, and that the model executes without fine-tuning or post-training — the extension to robotics of the shift from fine-tuned language models to prompted ones, in the company's framing (Introducing S1: In-Context Learning for Robotics).

The company reports a controlled comparison between demonstration prompting and the language prompting used by conventional vision-language-action policies, with both trained on identical data, architecture and compute across pre-training sets from 1,000 to 100,000 hours. On tasks drawn from the pre-training distribution, the two converge — 96% for the in-context policy against 89% for the language-prompted one at 100,000 hours. On tasks never seen in pre-training, they diverge: 66% against 9% at the same scale, a ratio the company describes as a sevenfold improvement. At the smallest scale tested the ordering reverses, with the language-conditioned policy at 53% against 43% for the in-context model, which Skild attributes to in-context models overfitting to a small amount of high-dimensional conditioning data (Introducing S1: In-Context Learning for Robotics).

Skild reports four long-horizon tasks absent from training — plant potting, pancake cooking, pour-over coffee and kit assembly — running up to ten minutes from a single demonstration, and logs eleven minutes between recording the potting demonstration and autonomous execution. It puts one in-context demonstration at roughly 380 post-training episodes, a crossing point estimated by interpolation, while noting that post-training eventually passes in-context performance, reaching 86% at 2,000 demonstrations against the in-context policy's 66% (Introducing S1: In-Context Learning for Robotics).

All of these figures are Skild's own, measured on two internal benchmark suites rather than a public benchmark, and reported as average per-step success rates under a protocol that uses human intervention to recover from failures so that every step is graded. None has been independently reproduced (Introducing S1: In-Context Learning for Robotics).

The post positions S1 against GEN-1.5 from Generalist AI, which it names as a concurrent approach and argues is "constrained to short horizons, in-distribution tasks, or both" — a characterization Generalist's own account of one-shot in-context learning does not share (Introducing S1: In-Context Learning for Robotics).

Infrastructure and partners

S1 is trained on Nvidia AI infrastructure, which the company credits with supplying the accelerated computing needed to train across its mixed robotics data (Introducing S1: In-Context Learning for Robotics). Nvidia also participated in the Series C alongside SoftBank, Macquarie Group and 1789 Capital (Source: techcrunch.com). The company states that S1 is deployed with commercial partners, without naming them (Introducing S1: In-Context Learning for Robotics).

Skild's data strategy is to collect from every available source rather than specialize. Its published comparison rates robot teleoperation highest on hardware proximity and lowest on scalability, egocentric video the reverse, with UMI hand-held collection and simulation in between; the company states it scales all four in house and spends three dollars on quality control for every dollar spent on collection (Introducing S1: In-Context Learning for Robotics).

Competitive position

Skild is one of several companies training large vision-language-action and embodied foundation models, a group that also includes Generalist AI, Figure AI, Physical Intelligence, Tesla and Google DeepMind, whose Gemini Robotics-ER 1.6 was released in April 2026 (Source: deepmind.google). The distinguishing claim in Skild's August 2026 release is the horizon length of the tasks learned from context, not the existence of in-context learning itself.

Relationships

• instance-of: AI Robotics — embodied AI and robot foundation models
• depends-on: In-Context Learning — the learning paradigm S1 is built around
• depends-on: Foundation Models — the pre-train-once, deploy-many pattern applied to robot control
• related: Generalist — the concurrent in-context robotics approach S1's release argues against
• related: S1 (Skild AI) — the company's flagship manipulation model
• related: GEN-1.5 — the competing one-shot embodied model
• related: Carnegie Mellon University — the institution both founders came from and remain affiliated with
• related: Figure AI, Google DeepMind — peer developers of robot foundation models
• related: Embodied Perception (Tacit Knowledge vs. Pattern Recognition), Multimodality — the perception and modality questions the model's video conditioning turns on

Open questions

• Whether S1's reported in-context results hold on any benchmark outside Skild's two internal suites; no independent evaluation has been published.
• Which commercial partners have S1 in deployment, and at what scale.
• Whether S1 and the Skild Brain are the same system under different names, or S1 is a manipulation-specific model within a broader family; the company's published material does not say.