Opt in to receive updates. Unsubscribe anytime.

Experience NEO

We're excited to announce Tom Sanocki as our new VP of Engineering. Tom joins the 1X team with a remarkable track record of building and scaling software organizations that have reached millions of people. He has operated across the full spectrum, from $30B+ public companies to seed-stage startups– across graphics, UGC platforms, VR/AR, and consumer products.


He was the first Engineering Director for Avatars at Roblox, led four engineering teams on Meta’s Horizon Worlds, and founded a VR interactive character platform that was acquired by Lytro. Earlier in his career, after earning his Computer Science Engineering degree from Princeton University, he contributed to Bungie’s Destiny 1 and several foundational Pixar films including Finding Nemo, Cars, Brave, and Up.


“The 1X team is solving the hardest problems for a truly general purpose consumer robot and doing it right now. Their commitment to safety and human values is exactly what we need, and I’m excited to join the team as we begin shipping the first humanoid robots to everyday people.”

Tom’s deep experience shipping complex, real-time, user-facing systems at massive scale — from UGC platforms and avatars to immersive VR worlds — will be instrumental as we extend NEO’s capabilities and bring the next generation of humanoid intelligence into the real world. His leadership will also play a key role in bringing 1X’s recently announced developer platform to life, empowering developers and creating a vibrant ecosystem of innovative applications for humanoid robots.


“Tom joining 1X is a milestone for the company. His expertise in graphics, real-time systems, machine learning, and character platforms is world-class, and exactly what we need to accelerate our software roadmap as we move from prototypes to products that leave the factory and land on doorsteps.”

Welcoming Tom Sanocki as VP of Engineering

Today, we're excited to announce that 1X has officially established the 1X World Model Lab

, our new frontier research organization dedicated to accelerating the path to fully autonomous humanoids through large-scale embodied world model pretraining. Also, we’re proud to welcome Sam Sinha– previously a founding research scientist at Luma AI and a leader in scaling multimodal generative video models– as our Founding AI Researcher and Head of World Models.

We put this team together on the heels of major breakthroughs with our 

. These advances enabled NEO to generalize to completely unseen tasks with zero-shot execution. For us, this was when the recipe for fully autonomous humanoids clicked:

The next frontier in AI requires scaling: web-scale media + egocentric human videos + sim + dexterous remote operated robot data + on-policy NEO data → real-world deployment for robot data collection and RL → abundance of data → physical AI

, scaling the fleet to support real world data collection is no longer the bottleneck, so we’re doubling down on increasing the pace of innovation of our world model to unlock fully autonomous humanoids as fast as humanly possible.


“World models will free humanoid robots from the trap of narrow tasks in static environments by unlocking generalization, adaptability, and self-learning. But to reach full autonomy fastest you must own the entire stack — down to pre-training your own video foundation models. That’s why we built the 1X World Model Lab.”

The 1X World Model Lab is a new embodied AI research organization focused on pretraining the foundation models to accelerate the emergence of embodied intelligence.

robotics is not a fine-tuning problem. To build truly general humanoids, we need to pretrain on the most important data from the very beginning.

Embodied intelligence requires models that see the physical world early, not just in post-training. This includes video, ego-centric human experience, interaction data, simulation, robot data collected with remote operation, and on-policy NEO data. Each provides a different signal about how the physical world works and how intelligent agents operate within.

The 1X World Model Lab will focus on large-scale AI pretraining, building frontier world models and reasoning models to enable spatial intelligence.

 These capabilities form the foundation for increasingly general and autonomous humanoid systems.

We believe the only durable moat in embodied AI will be data: the ability to collect, annotate, scale, and train on the right mixture of web-scale, human, simulated, remote operator-generated robot data, and robot-generated data.

1X is uniquely positioned to connect the entire learning loop—from pretraining and robot learning to evaluation, deployment, and the collection of on-policy data from real robots. By closing this loop, the 1X World Model Lab will accelerate the development of increasingly capable foundation models for humanoid robotics.

Welcoming Sam Sinha as Founding Researcher and Head of World Models

Sam brings world-class expertise in scaling large multimodal generative foundation models. At Luma AI, he played a key role in pushing the boundaries of image and video generation, including work on Uni-1. His deep experience in generative AI translates directly into building next-generation world models purpose-built for physical intelligence and humanoid robotics.

“We believe the only durable moat in AI is data — specifically, the ability to collect and learn from the optimal mixture at large scales. Embodied data has been treated as a second-class citizen for too long. Embodied AI is too important to be a fine-tuning problem. Scaling along every imaginable axis in robotics is the only way forward. 1X is uniquely positioned to close the full learning loop from data curation and pretraining, through real-world deployment.”

— Sam Sinha, Founding AI Researcher and Head of World Models

The 1X World Model Lab is building the team that will define the future of embodied AI. We are actively hiring exceptional talent across the three pillars of AI: data, infrastructure and learning.

 your background + proof of excellence and apply for roles on the 1X careers page.



The 1X World Model Lab | Est. 2026

Six months ago, we launched the preorder campaign for NEO The Home Robot.

NEO became one of the most talked about products on earth and shattered records for product launches. But most importantly, it created a community of early customers that trust us to deliver the world's first home humanoid to their doorstep.

It wasn’t just overwhelming, it was (and still is) deeply emotional for us. Some of the team here at 1X have been working on this dream for more than ten years with nothing but passion and deep conviction that the world needs helpful robots and that other people would love NEO as much as we do. After spending years being called crazy, it can be easy to doubt your conviction.

But on October 28th, we launched NEO to the world and booked out our entire production capacity for the next year in just 5 days (10,000 NEOs). With that, our mission became clearer than ever, and we’ve had our heads down for the last few months taking the right steps to deliver on our promise.

With that said, the biggest question we get to date is: “Where is my NEO?”

Our team has been doing their best to answer all of your emails and inquiries on X, but we figured it would be best to just answer directly.

First, it’s important to set a bit of context around the preorder and its position in the broader plan to bring NEO to homes everywhere.

 We have a product we are really excited about that we wanted to share with the world (and it’s already better than marketed. Breakthroughs like 

 and more to come will unlock even more exciting AI functionality than previously anticipated).

 We needed to figure out who wants a NEO, where they are in the world and how many people want a NEO to inform production planning and market entry plans.

 Since launching, we’ve been able to connect with dozens of early customers and learn about their needs– helping us hone in the feature-set we’re delivering.

So back to the question of “Where is my NEO”

In short, it’s coming. Some of you will get your NEO this year, some will get them later. We promised the first NEOs would ship in 2026, and we’re keeping that promise. We are working around the clock to get it to you, but want to make sure that what we ship does justice to the humanoid space (this product has been built on the shoulders of giants) and the trust you’ve placed in us.

There is a lot that goes into creating the first ever humanoid consumer product experience and we aren’t hiding from it. You have to take everything you’ve built, test it aggressively, iterate on it endlessly, and package it up to ship to customers that paid good money for a life-changing experience. There is a lot of trust in that exchange and we don’t take that lightly. The road ahead includes everything from scaling up manufacturing, expanding our internal home testing program (more 1X employees need NEOs!), and continuing to iterate on the product experience until it’s ready for the general audience. Starting today– we are going to be giving you more frequent updates on all fronts.

We have been making tremendous progress towards our milestones and today I couldn’t be more excited to take you inside our NEO Factory right here in Hayward, California, where we are producing more NEO’s than ever.

At 1X, we don’t just order parts from suppliers, assemble them, and call that manufacturing.

The NEO Factory in Hayward, California is America’s first vertically integrated high-volume humanoid robot factory. Spanning 58,000 sq ft and already employing 200+ team members, the NEO Factory has commenced full-scale production of NEO.

We design and manufacture NEO’s critical components in-house — including motors, batteries, structures, transmission systems, soft goods, sensors and more. From cutting metal parts to fully automated motor manufacturing lines that form copper coils, we maintain maximum control of our supply chain. This approach sets us apart from competitors that rely on foreign suppliers for critical subsystems–enabling not only faster iteration, but higher safety standards, efficient scaling, and resilience to geopolitical events that can impact supply chains.

With the NEO Factory in Hayward now online and our San Carlos facility coming online later this year, 1X has the capacity to produce up to 10,000 NEOs per year, with planned increases in automation expected to scale production toward 100,000+ units annually by the end of 2027.

Scaling NEO production was always part of the plan. But after the response to our preorder, we had to accelerate everything. With speed in our DNA, we stood up the NEO Factory in Hayward, California in just three months.

In the short time since receiving our final permits in January, we’ve hired more than 200 employees and have become fully operational.

Most people think humanoids are a robotics problem. They’re wrong. It’s a manufacturing problem. Production makes prototypes look easy.

So we built the machine that builds the machines:

Manufacturing Lines, Automations and Processes

 to create parts from scratch for New Product Introduction and mass manufacturing.

 at the core of all factory operations–built in-house to streamline NEO production. Instead of relying on fragmented tooling, it brings every stage of production into a single system that operates in real time.

designed to break things fast and surface failures–. 20M+ cycles and counting.

 lines to produce and test new builds in limited quantities before being introduced on the main line.

is where we assemble NEO’s joints and body parts for mass production.

 is where robots are fully assembled and stand for the first time before putting on covers and suits.

 is where semi-trucks pull up to the warehouse doors to drop off parts and materials.

We spend a lot of time talking about NEO, but hands down the best part about 1X are its people. While our content team was down in the factory building a story around the hard work everyone is doing down in Hayward to get you your robots– the stories of the factory were something they couldn’t leave without capturing.

To be clear, 1X is nothing without it’s people and we couldn’t be more grateful for everyone that is on this mission together

Nearly every component of NEO is designed and built from scratch at 1X using manufacturing processes and automations that we’ve been developing since founding the company in 2014.

Even for the parts we don’t produce in-house at the NEO Factory, we work directly inside our suppliers’ facilities to design, develop, and enforce our own quality standards. This level of true vertical integration allows us to deliver the lowest-cost hardware, iterate at an industry-leading pace, and protect our supply chain from geopolitical risks.

Today, we’re excited to take you inside the process, from raw materials to finished robots walking off the line.

Since the start of production in Hayward, our team has manufactured 17,000 motors.

We’ve been perfecting the Revo2 motor– the heart of NEO’s tendon-driven actuation system– since 2015, beginning in Moss, Norway. Producing motors in-house allows us to push the physical limits of torque-to-weight density.

Our motor manufacturing starts with a spool of copper and ends with a fully assembled tendon-driven actuator. Every step is automated by machines designed and built by 1X engineers– from winding our own copper coils to fabricating stators from electric steel. The result is the world’s highest-performance, lowest cost electric motor for humanoid robotics.

Our state-of-the-art motor electronics remove the traditional constraints of motor design, allowing us to design net new motors with dramatically higher torque density.

By controlling every layer– from electric steel selection to manufacturing methods, software, and custom electronics– we can extract maximum performance from the Revo2 motors. This level of vertical integration, engineered from first principles, is what enables us to co-develop components that work together at a level simply not possible when relying on off-the-shelf parts.

NEO’s tendons are the result of more than a decade of innovation in materials science and advanced manufacturing processes. They are engineered to handle heavy payloads while delivering safe, energy-efficient, and remarkably quiet movement– all with industry-leading durability and silent operation.

Every tendon is produced using our proprietary process, which begins with careful material selection, followed by precision braiding and specialized post-processing. This ensures each tendon is perfectly prepared for integration into NEO’s Tendon Drive assembly.

We manufacture NEO’s 22 DoF hand (more on those soon) on our dedicated production line.

Every hand begins with material selection for tendons and rigorous in-house testing. From there, we assemble the full tendon system, integrate our custom electronics, and install the 1X Motors to drive each degree of freedom.

The structural components– fingers, palms, wrists, and forearms– are manufactured directly on the line. Finally, we mold the soft “flesh” and outer gloves using our proprietary polymers, seamlessly integrating an advanced tactile sensing stack throughout.

This end-to-end in-house process gives us complete control over quality, performance, and iteration speed for one of the most important and complex parts of NEO.

The 1X Battery Pack is manufactured with aerospace-grade cells, allowing us to achieve performance and safety levels that are simply not possible with off-the-shelf batteries.

Our proprietary Battery Management System (BMS) samples at 100 Hz, which is roughly 100 times faster than typical commercial packs. This enables real-time monitoring of battery state of health and highly accurate predictions of future performance, state of charge, and state of function. The result is a safer, longer-lasting battery with exceptional fast-charging capability.

We designed the entire battery manufacturing line for high-volume automation using low-cost, highly reliable components. It can complete a weld every 500 milliseconds, giving us the capacity to produce 10,000 batteries per year. A key innovation to 1X Battery Pack production includes innovative side-cooling on both the top and bottom of the battery pack. This approach delivers superior thermal management while supporting rapid, scalable production.

The outcome is a battery that combines industry-leading safety, reliability, fast-charging, and longevity at the lowest possible cost.

Every NEO rolling off the production line comes equipped with the latest 

 embedded in the NEO Cortex. This integration enables powerful real-time AI inference directly on the robot for safety-critical functions, perception, reasoning, and decision-making—without heavy cloud computing.

NEO Cortex, AKA NEO’s brain, tightly integrates on-board compute with custom sensors including dual 8.85MP 90Hz stereo fisheye cameras, beam-forming microphones for directional audio, IMUs and more. Each component is mounted on a custom racking engineered and built in-house, paired with a thermal management system that keeps powerful hardware running efficiently and whisper-quiet.

The first NEO took 1 year to go from CAD to walking its first steps. The latest version, produced in our Hayward Facility, went from CAD to walking in just 1 month.

This acceleration is powered by our NPI line co-located near the Machine Park, where hardware design and manufacturing engineers can go from CAD release to physical parts in as little as a few hours. These rapid prototypes allow us to test new designs, identify issues, and implement solutions at an unprecedented pace– all before new designs are released to mass production.

At the NEO Factory in Hayward, California, we designed the assembly line around continuous flow rather than traditional departments. The entire system eliminates dead zones and queues, favoring single-piece flow wherever possible and batching only where physics demands it.

Each station receives pre-kitted modules, removing the need for searching and guesswork. Functionality is added through progressive bring-up, with in-line verification at every critical step. This ensures issues are caught and resolved immediately, rather than at the end of line.

The line is further strengthened by intelligent tooling that enforces correctness by design: if the part fits, it’s right. Cycle times are tightly balanced across stations to prevent hidden bottlenecks, while proprietary tendon joint conditioning eliminates early-life failures before final integration. Every build is fully digitally traceable down to the individual component and torque value.

The result is a tightly controlled process where quality and throughput scale together.

As we scale robot production, automation is critical to delivering both the volume and uncompromising quality our targets demand. We prioritize highly repeatable processes with minimal variation— tasks where consistency, reliability, or cycle time matter most. Whether that means holding micron-level tolerances or executing uniform process steps at scale, automation eliminates human-induced error and ensures every unit meets the same standard.

Over the past decade, we’ve developed proprietary assembly automations that now handle critical steps across the build —from precise stator insertion into the housing to controlled soldering. Each system is engineered specifically for its task, optimizing alignment, force application, thermal conditions, or motion profiles as needed. Looking ahead, our production lines will integrate a hybrid approach: proven industrial automation for speed and repeatability, combined with humanoid robots for greater flexibility. This combination will let us scale output, adapt quickly to design changes, and continuously improve efficiency, consistency, and overall factory throughput.

Today, our robots have already started to help around the factory. They are actively collecting real-world data and using the 1X World Model to learn practical tasks– from stocking parts for our assembly technicians to performing basic warehousing and logistics operations.

In the coming months, they will take on broader roles such as facility security and other longer-horizon tasks where reliable, helpful hands are needed most.

Long term, the possibilities are near limitless. We are incredibly excited to share regular updates as NEO evolves from factory assistant to a true general-purpose robot– one that will eventually handle everything from household chores to building robots, supporting chip fabs, powering data centers, and more.

NEO Factory | Building Your NEO

We’re excited to announce Amy Lentz as our new SVP of People at 1X.

She joins 1X after a distinguished career spanning the Obama White House, Teach For America, and TOMS, where she most recently led organization-wide transformation as Chief People Officer. She has built her career at the intersection of strategy, operations, and human performance, making the case that people strategy is a true competitive advantage, not a cost center.

“Joining 1X feels like the perfect next chapter. The opportunity to help shape a company that is building advanced humanoids with genuine care for human values, safety, and meaningful coexistence is incredibly inspiring. I’m excited to bring my experience to a team that treats people strategy as core to long-term success.” - Amy Lentz

Amy has advised ambitious leadership teams across tech, including Perplexity AI, and has built an audience of over one million professionals through her platform Hack Your HR, where she turns executive behavior into practical, actionable frameworks. Her core thesis, that emotional regulation is your corporate superpower, reflects the kind of thinking that will help 1X continue building a world-class organization.

“Amy’s leadership is exactly what we need as we scale. She brings deep expertise, a sharp strategic mindset, and an unwavering belief that investing in people is how great companies win. At 1X, our people are everything, and Amy’s arrival is a powerful testament to how seriously we take our team as we work to bring NEO to the world.” - Bernt Børnich

Welcoming Amy Lentz as SVP of People

Training humanoid robots requires a massive amount of infrastructure behind the scenes — high-fidelity simulation environments, large-scale model training, synthetic data generation, and powerful onboard compute for real-time inference.

At 1X, we use several parts of NVIDIA’s robotics platform throughout our development pipeline, from generating simulated training data to running models directly onboard our robots. With NVIDIA’s new open source collection of physical AI agent tools announced this week at GTC Taipei, we’re excited to share a bit more about how these tools fit into our stack and will further accelerate our workflow.

This is a big step forward for the entire industry, and we’re proud to be one of the robotics companies already leveraging NVIDIA’s agent-ready physical AI stack to develop more capable humanoid robots faster.

NVIDIA Isaac Sim and Isaac Lab: Simulation and robot learning

Before a robot learns a new capability in the real world, it often practices thousands or even millions of times in simulation. This allows us to explore new behaviors and generate training data before deploying models on physical robots.

, open-source frameworks for robot simulation and learning, to generate photorealistic environments where our robots can train and evaluate tasks. By training our world models in these physics-based simulation environments, we can run task executions, measure whether a model successfully completes them, and generate additional training data to improve performance.

large numbers of GPU simulations in parallel

 in the cloud. We use this to train reinforcement learning controllers for locomotion and other foundational robot capabilities. Running these simulations at scale exposes the models to far more scenarios than would be possible with physical robots alone.

Cloud Computing: Training models in the cloud

Training embodied AI models requires significant computation.

We use NVIDIA Blackwell HGX B200 GPUs to train our proprietary robot foundation models across both simulated and real-world robot data. These models learn to connect perception with action — interpreting visual scenes and translating them into physical behavior.

Access to large GPU clusters allows us to iterate quickly during both mid-training and post-training stages as we improve robustness and generalization across environments.

NVIDIA Jetson Thor: Running models onboard the robot

Once models are trained, they need to run directly on the robot with minimal latency. Real-world interaction requires fast perception and decision-making while operating within strict power and size constraints– 

 is the only product on the market that is built to support NEO’s requirements for onboard compute.

Thor allows us to run large neural networks locally on the robot itself, supporting real-time perception, reasoning, and control with optimized inference performance.

Running multimodal AI models directly onboard 

 allows the system to respond quickly to its environment while minimizing reliance on external computers.

Vision is one of the primary ways humanoid robots understand the world around them.

As part of our perception stack, we use NVIDIA

Argus to process camera data and support low-latency visual inference. This helps maintain a tight perception-to-action loop when the robot is navigating environments or interacting with objects.

Together, these tools support the full lifecycle of robot learning — from simulation, to large-scale training, to real-world deployment.

Combining this infrastructure with our robotics and AI systems allows us to move efficiently from simulated learning to physical capability as we continue developing NEO and advancing our humanoid robotics platform.

Developing NEO’s AI Using NVIDIA’s  Robotics Platform 

Home robots need common sense behavior and a deep understanding of the physical world.

Many robot foundation models today are vision-language-action models (VLAs), which take a pretrained VLM and add an output head to predict robot actions (

). VLMs benefit from internet-scale knowledge, but are trained on objectives that emphasize visual and semantic understanding over prediction of physical dynamics. Tens of thousands of hours of costly robot data are needed to teach a model how to solve tasks considered simple for a human. Additionally, auxiliary objectives are often used to further coax spatial reasoning of physical interactions (

In this blog, we introduce our video-pretrained world model, 1XWM, integrated into NEO as a robot policy. While VLAs directly predict action trajectories from static image-language input, our world-model based policy derives robot actions from text-conditioned video generation. By leveraging the world dynamics inherent in internet-scale video, our world model generalizes to novel objects, motions, and tasks without pre-training on large-scale robot data or any related teleoperated demonstrations. This represents a shift towards a new regime: allowing robot intelligence to benefit from the scaling of video pretraining, enabled by a hardware stack designed for high fidelity transfer from human embodiment.

Internet video implicitly encodes the structural priors of reality: how people and objects move, where forces are applied, and the implicit constraints of interaction. Accurate translation of video into action comes from more than just a model—it benefits from an embodiment that is kinematically and dynamically congruent with the human form.

The hardware as a first-class citizen in the AI stack closes the human-robot translation gap. By combining embodiment with human-like compliance, interaction dynamics—friction, inertia, and contact behavior—often match human motion closely enough for the model’s learned priors to remain in distribution. What the model can visualize, NEO usually can do.

Today, frontier text-to-video models like 

 generate strikingly realistic videos. However, zero-shot generations from these models are not grounded in robot embodiments and often fall short along key axes required for control:

: Is the rollout consistent with the robot’s camera intrinsics and egocentric viewpoint? Does it preserve depth and the precise spatial relationships required for manipulation?

: Does the robot in the rollout move in ways feasible for the embodiment, respecting morphology, joint limits, velocities, and actuation constraints?

: Does the rollout avoid physically impossible outcomes (e.g. object teleportation) which would not translate to real-world success?

. To bring video knowledge to world model form, we utilize a two-stage grounding process enabled by our integrated stack, following existing works like 

 A text-conditioned diffusion model trained on web-scale video, mid-trained on egocentric human data, and fine-tuned on NEO-specific sensorimotor logs. It predicts the high-fidelity evolution of a scene with strong visual, spatial, and physical accuracy.

 We bridge pixels to actuators by training an IDM to predict the exact action sequence required to transition between the model's generated frames. We use the IDM’s metrics and rejection sampling to enforce kinematic correctness of generations.

At inference time, the system receives a text prompt and a starting frame. The WM rolls out the intended future, the IDM extracts the necessary trajectory, and the robot executes the sequence in the real world.

At test-time, we send a prompt from our frontend to the model inference server, which will then execute the action on-robot.

The 1XWM backbone is built upon a 14B generative video model. To adapt this model to NEO’s embodiment, we use a multi-stage training strategy:

 We train on 900 hours of egocentric human video to align the model with first-person manipulation tasks. At this stage, the model captures general manipulation behaviors but struggles to generate videos of NEO performing tasks.

 We then fine-tune on 70 hours of robot data, which adapts the model to NEO’s visual appearance and kinematics.

 show that the prompt adherence of visual foundation models can be significantly improved by training on descriptive visual captions. However, many egocentric datasets contain only brief task descriptions. To address this, we prompt a VLM to generate more detailed captions to use for training through a 

To select the best 1XWM backbone checkpoint, we use an evaluation metric that compares the 

 distance between the ground truth future action sequence and the actions from our trained IDM run on the generated video. This helps filter for generations that not only look good visually but also yield accurate actions.

Similarly to DreamGen, we use a two-image IDM predictor with a sliding window of W=8 frames for efficient training and inference. However, we use a 

 backbone with a separate flow matching head instead of a single diffusion transformer. Frames at times t and t+W are passed through the depth backbone and the embeddings are used as conditioning for the flow matching head. The IDM is trained on 400 hours of unfiltered robot data including random play data and motions that don’t correspond to any meaningful task. This allows us to faithfully track NEO’s motions anywhere.

At test-time, given the starting frame and a text prompt instructing NEO on what to do, 1XWM rolls out future video frames. We then extract the corresponding robot action trajectory with the IDM, and execute it directly on the robot. To ensure smooth trajectories, the IDM’s outputs are timewise averaged across a batch of initial noise values and sliding windows.

Caption: Our NEO post-training dataset contains primarily high quality pick-and-place data (98.5%), filtered for table top manipulation with hands in view. By leveraging the web-scale pretraining of the base video model, 1XWM can generalize to a wide range of unseen objects, environments, and tasks.

We first seek to understand: what are the limits on task generalization beyond what NEO has already seen and done? How closely does generated video align with real-world execution?

manipulate unseen objects with interesting affordances

perform entirely novel tasks requiring new motions

We observe that 1XWM generated videos generally align well with real-world execution. Generated and post-factual video can look very similar viewed side-by-side, showing 1XWM has strong spatial, kinematic, and physical understanding.

Grasping

New Behaviors


Next, we try tasks requiring two-handed coordination and human interaction—abilities not included in our training dataset. This suggests that such knowledge comes from video pre-training and ego human mid-training. Because NEO’s embodiment is so similar to humans, the affordances learned in human video data translate directly.

Two-Handed

Human Interaction


Sometimes, generations tend to be overly optimistic about task completion and depth understanding. Generated rollouts can look visually plausible while subtly violating real-world constraints (e.g. object consistency, depth, geometry, contact). Our post-training recipe significantly reduces these errors, but monocular pretraining can still lead to weak 3D grounding, where the real robot undershoots or overshoots even when the generated video “succeeds.” This opens future work to integrate depth or stereo for better spatial grounding. 

Beyond anecdotal examples, we show real-world results measuring the performance of 1XWM on in-distribution (ID) and out of distribution (OOD) tasks, running 30 times each. 1XWM succeeds with stable success rates across diverse action primitives, although some dextrous tasks remain challenging (e.g. pouring, drawing).

Can we connect video quality and task success?

If so, we can measure and improve video quality using visual metrics and estimate the likelihood of real-world success. 

Sometimes it is visually obvious whether a generated rollout is likely to succeed. For example, prompting 1XWM with “pull tissue” can occasionally generate videos of NEO picking up the tissue box instead. We’ve generally found close to 0% success rate when executing bad generations. 

 can improve task success. Inspired by this, we try generating multiple rollouts in parallel and executing the best one. We study the “pull tissue” task between one and eight parallel generations, finding that the ability to select the highest quality generation from eight choices does lead to improved task success.

This selection process can be done manually, but is amenable to automation with a VLM evaluator. We leave exploration of best-of-N sampling strategies to future work. For simplicity and disambiguation with our quality of video ranking, all other results in this blog post use a single generation per attempt.

The importance of egocentric data and high-quality captions

Given our hypothesis of correlation between video quality and task success, we visually ablate a few training choices, specifically the impact of upsampling captions and training on egocentric human data.

To evaluate, we generate videos of NEO performing tasks and then ask for human feedback. We use three evaluation datasets, each containing 500 starting image-prompt pairs:

 contains hard tasks and backgrounds in-distribution with our robot training data: mostly pick and place in scenes with clutter and challenging object positions.

 consist of a split across novel tasks: whisk the bowl, pull tissue, relative size object identification (pick bigger object), bimanual manipulation, etc. collected with simple backgrounds in the real world.

 is composed entirely of pick tasks, using seed frames generated by a text-to-image (T2I) model, which randomly samples out-of-distribution household objects and backgrounds.

In-distribution

New Tasks

We ask human annotators to review each generated video and accept or reject it based on physical plausibility, task completion, and consistency with NEO’s embodiment and capabilities.


We find that upsampling captions improves video generation quality on every evaluation split. Upsampled captions better match the detailed text conditioning used during video-model pretraining and also provide clearer conditioning for task-specific motion.

Adding egocentric human data improves generation quality on both 

 splits. This is consistent with our hypothesis that ego human data contributes a transferable prior for manipulation that maps well onto NEO’s humanoid embodiment. For in-distribution tasks for which we have good NEO data coverage, egocentric data may dilute the post-training mix and cause a neutral or negative change in visual score.


Finally, we run real-world ablations on egocentric mid-training and caption upsampling, techniques we found to improve visual generation quality. For each task, we run all models 30 times in the same setting, on the same robot, and without test-time filtering. We do not control further with blind evaluation or checkpoint selection.

, we find that all models perform similarly, suggesting that in-distribution tasks are least sensitive to changes in data or caption.

 task was the most challenging across models. Models without either the egocentric human midtraining phase or caption upsampling incorrectly interpret the instruction. We find that our model with both is the only model to achieve a nonzero success rate.

As a whole, we do see a connection where including changes that impact visual model quality improve task success in real-world experiments.

Ego human data

Only NEO data

NEO + ego human data

1XWM transfers the knowledge of web-scale video data to imagine and then execute diverse tasks on NEO. From here, we’re excited to improve 1XWM to solve more complex, longer-horizon tasks required for useful household autonomy.

Building on the rapid progress observed in generative video models and equipped with data from our NEO production ramp, we will continue scaling 1XWM to improve quality and task success. Today, the 1XWM backbone takes 11 seconds to run, using multi-gpu inference built in collaboration with our friends at Verda (Antonio and Aditya). Each inference generates 5 seconds of real-time video, and the IDM takes 1 second to extract actions from the generated rollout. To solve reactive tasks and contact-rich manipulation, faster inference will improve reaction time, or latency between prompt and action execution. As we consider tasks longer than 5 seconds, closed-loop replanning with memory context will help us handle drift, partial observability, and recovery.

We don’t need perfect performance across the board to make deep progress. Nonzero success across a broad set of tasks provides the lever to unlock self-improvement.

1XWM creates a flywheel where exploration, evaluation, and policy refinement are driven by NEO’s own experience, rather than being limited by expert demonstrations. When learning is driven by experience, NEO can only improve from here, and we are excited to build a future where NEO can teach itself to master any task in any home.

If this sounds motivating to you, come help us build the future: 

https://1x.recruitee.com/o/ai-research-engineer-world-models

1X World Model | From Video to Action: A New Way Robots Learn

Palo Alto, CA, October 28, 2025 – 1X is proud to announce the launch of NEO, the world’s first consumer-ready humanoid robot designed to transform life at home. NEO automates everyday chores and offers personalized assistance so people can spend more time on the things that matter.

With intuitive, user-friendly features and a revolutionary hardware platform, NEO is the safest, most affordable, and most capable humanoid on earth. NEO brings the science-fiction idea of robots from the movie screen into reality — creating space for us to enjoy our time at home, instead of coming home just to do more work.

"Humanoids were long a thing of sci-fi... then they were a thing of research, but today — with the launch of NEO — humanoid robots become a product. Something that you and me can reach out and touch. NEO closes the gap between our imaginations and the world we live in, to the point where we can actually ask a humanoid robot for help, and help is granted."

- Bernt Børnich, CEO and Founder, 1X

The Chores feature allows owners to give NEO a list of tasks to complete around the home, schedule a time for when they want them completed, and come back to a cleaner space every day. NEO can also handle chores in real time while you relax or work nearby. With the click of a button or a simple verbal command, NEO transforms into a personal housekeeper; tackling tasks like folding laundry, organizing shelves, and tidying spaces. For any chore that NEO doesn't yet know, owners can schedule a 1X Expert to guide it through unknown tasks — helping NEO learn while getting the job done.


NEO is built to engage naturally while providing help and utility through conversation. With a built-in large language model (LLM), NEO gives owners instant access to knowledge and personalized assistance without relying on devices with screens.

Its Audio Intelligence allows NEO to recognize when it is being addressed, ensuring it listens and responds only when appropriate. Visual Intelligence adds contextual awareness — letting NEO use what it sees to enhance interactions, like recognizing ingredients on a kitchen counter and suggesting dishes to make. Memory enables continuity across conversations, allowing NEO to recall past context and adapt its responses over time. From scheduling appointments and setting reminders for birthdays to keeping grocery lists and tracking progress in language lessons, not only does NEO free up time, but also valuable mental space.

NEO Operates Autonomously and Learns New Skills Over Time

The core of the NEO experience is driven by AI. Each NEO will be able to have conversations, and navigate around the home to provide help where needed. NEO will arrive to early owners with the ability to handle basic tasks, and with continued use, learn and develop new skills. From the first day, NEO performs functions like opening doors for guests, fetching items, and turning off the lights at night — and grows in abilities with every software update. Leveraging 1X’s AI model, which is trained on real-world data, NEO is able to adapt to the world’s most diverse environment: the home.

At the core of the NEO experience is a revolutionary hardware platform that succeeds 1X’s prior models (NEO Gamma and NEO Beta). NEO is built with 1X’s patented Tendon Drive that uses the highest-torque density motors on earth to drive its tendon-based transmissions. This one-of-a-kind actuation system creates gentle movements that make NEO uniquely safe and compliant around people.

NEO represents a generational leap in humanoid hardware with the addition of Human Level Dexterity (22 DoF Hands) and a head-to-toe soft body made of custom 3D lattice polymer structures. At just 66 pounds (29.94 kg), NEO is capable of lifting over 150 pounds (68 kg) and carrying 55 pounds (24.95kg), and has a noise level of 22dB — making it more silent than a modern day refrigerator. Additionally, NEO possesses built-in communication with WiFi, Bluetooth, and 5G as well as a three-stage speaker in the pelvis and chest area, giving owners a built-in mobile home entertainment system.

NEO is approachable and comfortable to be around. NEO’s essentialist design aims to complement living spaces rather than disrupt them – with organic neutral tones, and a soft knit suit and shoes.

NEO is now available for pre-order, with first orders shipping to consumer homes in 2026.

NEO is now available via 1X’s online store [

Welcoming Tom Sanocki as VP of Engineering

Discover

The 1X World Model Lab | Est. 2026

NEO Factory | Building Your NEO

Welcoming Amy Lentz as SVP of People

ExperienceNEO