Coming out of CES 2026 and its unprecedented robotics display, expectations are through the roof. So is skepticism.
As critics question whether the so-called ChatGPT moment for general-purpose robots is really imminent, tech leaders and their powerful friends remain bullish on a robot-run future. Perceived leaders in the global humanoid race kicked off the year with a flurry of claimed breakthroughs. Whether they’re major enough to win over doubters is an open question.
1X NEO’s Self Learning
Silicon Valley-based 1X Technologies just announced what it describes as a major architectural breakthrough in robot intelligence.
“This opens a new path for robotics learning, one where robots learn by teaching themselves using the data they generated on their own,” 1X Founder/CEO Bernt Bornich said in a video announcement.
1X, which was founded in 2014 in Norway and is now headquartered in Palo Alto, says its self-developed world model empowers its humanoid NEO to carry out tasks it’s never been explicitly trained to perform. On Bloomberg Technology, 1X’s CEO said: “Now all you need is the robots teaching themselves how to do all these tasks by actually experimenting and doing this in the real world.”
The 1X World Model is built on a very large AI system that learns by imagining short video clips of what is likely to happen next. It’s trained using hundreds of hours of first-person human video to understand everyday hand movements and object interaction and then refined with a smaller amount of real robot data to match NEO’s body in motions. A separate control model turns the predicted movements into actual motor commands.
Together, the system imagines a few seconds into the future before the robot acts. According to 1X, this means NEO can handle new objects and tasks without being specifically programmed for them. NEO basically generates short clips as it decides what to do next. Creating the AI-generated prediction videos requires powerful computing, so it’s largely run on external systems rather than the robot’s hardware.
The core model is pre-trained on large web-scale datasets similar to those used for modern generative video models like Sora Runway. The footage shows people and objects interacting in everyday situations. Next, the model is trained on egocentric human video featuring people performing tasks. This helps the robot brain understand human behaviors in a way that closely matches how a human experiences the world.
In tests, the imagined videos produced by the model closely matched what wound up happening in the real world. Using the system, NEO had the most success with simple, well-defined object interactions. It has a close to 95 percent success rate steaming a shirt, 80 percent grabbing chips, and 75 percent opening a sliding door. The success rate dropped as tasks became more physically complicated and contact-rich.
1X is preparing to ship its NEO humanoids to early adopters. The robots, which are priced at $20,000 to buy outright or $500 a month to rent, will collect data from those home deployments to improve the world model. 1X plans to rely on offsite teleoperators for initial deployments for tasks NEO can’t handle independently but the company says it could go full auto by 2027.
Skild AI’s $14B Robot Brain
Another well-funded US startup, Pittsburgh-based Skild AI, is also leaning into human video as it builds a universal brain for all robots.
The startup, launched in 2023 by Carnegie Melon professors, just shared footage of a humanoid and quadrupedal robot duo cleaning its office to demonstrate its progress. In a blog post, Skild researchers pointed to the vast amount of existing human-centric video from sources like personal wearable cameras to online instructional content.
“If we look at biological intelligence, the solution is hiding in plain sight,” the researchers wrote. “Humans don’t learn to make tea by being told the exact Newton-meters of force to apply to a kettle. We learn through observational learning. We possess a foundation of kinematics and dynamics that allows us to watch a visual demonstration, internalize the intent, and map those actions onto our own bodies.”
However, the team said using video as the primary training signal brings major technical challenges. Videos don’t directly capture physical information like force, torque, and tactile feedback, and robots differ in form and function from people. Skild says its so-called omni-body learning solution bridges the embodiment gap. Rather than replicating precise movements, the system focuses on understand intent and mapping it onto different mechanical hardware.
Skild is pursuing a general-purpose robotics brain with a sizable war chest. The startup just announced closing a $1.4 billion Series C funding round led by Japanese tech giant SoftBank. It’s now valued at more than $14 billion. Skild says it plans to massively scale data and put robots into real-world settings like warehouses, factories, construction sites, datacenters, and logistics hubs to accelerate progress with the funds.
LimX’s COSA
Meanwhile in China, the Shenzhen startup LimX Dynamics is boasting what it calls the world’s first agentic operating system for humanoids.
According to LimX, its Cognitive OS of Agents (COSA) system lets robots function as autonomous beings that can perceive, reason, remembers, and act continuously in the real world. The startup, launched in 2022, demonstrated the system with its flagship humanoid robot Oli.
The OS integrates high-level reasoning for the robot to interpret goals, proactive memory for contextual awareness, and a whole-body control foundation model that governs movement, balance, and manipulation. LimX says the architecture treats motion as part of cognition itself rather than a mechanical afterthought.
Mentee’s AI First Approach
Another startup founded in 2022, Israel’s Mentee Robotics, is also showing off the AI-first capabilities of its flagship humanoid.
In a recent demo, the company showcases its flagship robot, MenteeBot, learning to change a fellow machine’s battery from a single human demonstration. The 3-minute video shows the humanoid observing a person performing the battery replacement task from its own point of view. The demo is captured as raw video and converted into structured representations of full body motion data, object interactions, and task sequencing.
The AI reconstructs the tasks inside a simulation where the humanoids digital twin trains across thousands of variations to account for real-world uncertainty and edge cases. Once training is complete, the learn behaviors are transferred to the hardware through a process called sim tore real, short for simulation to reality. The result is a robot that can pull off the task independently and potentially teach it to other robots just like a human apprentice.
Mentee is scaling production of its third-generation bot as it transitions from startup to a subsidiary. Mobileye says it’s entered a definitive agreement to acquire mentee for about $612 million in cash and up to 26.2 million shares of its common stock. The total cash value is reported at around $900 million.