HomeFull speed ahead. Keep track of advancements in every frontier
Skild BrainSkild AI—Skild AI—Skild BrainSkild AI—Skild BrainSkild AI—Skild BrainSkild AI—Skild BrainSkild AI—
Skild Brain
Courtesy of Skild AI · Press kit, editorial use (opens skild.ai)

Skild Brain

by Skild AIUS

DeployStage 4 of 5

S1 model released Aug 2026. Skild reports $100M+ in annual recurring revenue from 60+ paying customers.

Updated 30d agoChecked 25 Sep0 updates this week

Milestones

Next · Commercial deployments with robot makers
  1. Skild Brain unveiledJul 2025Complete.
  2. Commercial deployments with robot makersNowCurrent milestone.

Most important updates

  • 30d ago
  • 18 Aug 2026
  • 15 Apr 2026
  • 19 Mar 2026
  • Jul 2025

Current obstacles

  • One model, many bodiesRobots differ in joints, sensors and weight, so one shared model often does worse than one built for a single robot.

Physics limits

  • Every body has different physicsMass, reach, joint limits and motor strength differ between robots. A shared model must infer them from a few moments of motion; if it guesses wrong, it pushes too hard or falls.
  • Video shows motion, not forceHuman videos reveal where hands go but not how hard they grip or push. Force must still be learned from robot experience, which is slow and costly to collect.
  • Errors compound over long tasksA 20-step job at 95% success per step finishes only about 36% of the time. Long tasks need near-perfect steps or reliable recovery.

How it works

3 parts
One of the many robot bodies Skild Brain drives: a pair of arms placing dishes in a rack, using the same model as its legged robots
One of the many robot bodies Skild Brain drives: a pair of arms placing dishes in a rack, using the same model as its legged robotsCourtesy of Skild AI · Press kit, editorial use (opens skild.ai)
Many bodies

Trained across robot shapes

The model practises in simulation on many different robot bodies, so its skills do not depend on one robot's size, legs or arms.

Human video

Learning by watching people

Besides robot data, it learns from videos of humans doing tasks, a far larger source than teleoperated robot recordings.

S1

One demo, new task

Skild's S1 model (Aug 2026) watches a single video demonstration of a new task and performs it without retraining.

Update log

7 updates

Thu 10 Sep

  • Major: BlogCommercial

Tue 18 Aug

  • Major: BlogSoftware

Wed 15 Apr

  • Major: BlogCommercial

Thu 19 Mar

  • Major: BlogPartnership

Mon 12 Jan

  • Minor: BlogResearch

Wed 24 Sep 2025

  • Minor: BlogResearch

Jul 2025

  • Major: BlogSoftware

About Skild AI

The team behind Skild Brain

Skild AI

Skild Brain, one AI for many robots

Skild AI, founded in Pittsburgh in 2023 by Carnegie Mellon professors Deepak Pathak and Abhinav Gupta, builds the Skild Brain, one AI meant to control many robot types. A SoftBank-led $1.4B round in January 2026 valued it above $14B.

  • Valuation went from $1.5B (2024) to $4.5B (2025) to over $14B (Jan 2026).
  • Pitches one 'brain' for robot dogs, humanoids and arms, trained largely in simulation and on human video.
  • SoftBank leads its funding and, through ABB Robotics, could help it reach factories.
Founded
20233 yrs
Headquarters
United States
Status
Private
Valuation
$14BprivateJan 2026
Raised
$1.7B2 rounds
Last round
Series C · $1.4B14 Jan 2026$14B post
Works in
RoboticsRobot brainsAI
Coverage
1 program · 8 updateslatest 30d agochecked 25 Sep
Lead investors
SoftBankLightspeedCoatueBezos Expeditions
Primer