Genesis AI raises $500 million for robots built for fine manipulation
Genesis AI has raised a $500 million Series A to build robots for the fiddly work industrial automation still struggles with: folding cardboard, peeling adhesive labels, sorting objects, and manipulating wires. Khosla Ventures, Premji Invest, and Ecl...
Genesis AI raises $500 million, with robot dexterity still facing the factory test
Genesis AI has raised a $500 million Series A to build robots for the fiddly work industrial automation still struggles with: folding cardboard, peeling adhesive labels, sorting objects, and manipulating wires. Khosla Ventures, Premji Invest, and Eclipse led the round, according to TechCrunch’s report published November 17, 2026.
The company combines human demonstrations, force-sensing gloves, simulation, a generative AI model, and its own humanoid hardware. It wants to control development from training data through to the machine doing the work.
CEO Zhou Xian says Genesis has “cracked dexterous manipulation.” The demonstration TechCrunch observed supports a narrower claim. Two off-the-shelf robotic arms completed a multistep packing workflow, an impressive sequence of operations. On a second run, the system fumbled a box and struggled to peel a small adhesive label.
Reliable repetition across changing conditions and long shifts remains an unresolved engineering problem.
Why the gloves matter
Genesis collects training data from more than 100 workers around the world wearing sensor-equipped gloves and headset cameras. Its robots wear the same gloves, and their hands approximate human hand proportions.
This setup addresses a persistent difficulty in learning from human demonstrations: people and robots don’t experience a task in the same way. A camera can capture where a worker’s hand moves, but provides much less direct information about grip pressure, fingertip contact, or changing resistance as the worker pulls an adhesive label. Force measurements capture some of that missing information.
The packing demonstration shows why it matters. Picking up a cardboard box requires enough grip to prevent slipping without crushing it. Folding a flap changes the contact geometry. Peeling a label requires separating thin material from its backing, then controlling something that can stick to the wrong surface. Each step calls for different contact behavior.
Human-shaped hands and shared sensing equipment could make those demonstrations easier to translate into robotic actions. Similar proportions reduce some differences in how fingers reach and grasp objects. Using the same sensors may also help align observations collected during human work with those collected during robot operation.
The rationale is plausible, though human and robotic hands still differ in actuation, compliance, movement limits, and wear. The report doesn’t disclose sensor specifications, sampling rates, control-loop timing, or how Genesis maps demonstrations into action targets. Engineers would need those details to assess how much performance comes from the learned model and how much depends on the surrounding control system.
Owning the data, model, and machine
Founded in 2024 by Xian and president Théophile Gervet, who met as graduate students in a Carnegie Mellon University lab, Genesis announced a $105 million seed round in July 2025. The new financing supports its vertically integrated strategy: generate most of its own training data, develop its own model, and design the robot alongside both.
That gives the engineering team room to fix problems across the system. It can change a hand’s geometry, modify sensors to capture missing signals, or collect demonstrations of a task that keeps failing without waiting for separate vendors to coordinate.
It also gives Genesis a lot to manage: data collection, model training, simulation, mechanical design, manufacturing, and field maintenance. Improvements in one area can leave another as the bottleneck. An actuator’s service life remains a constraint even when the manipulation software improves.
Simulation is a major part of the approach. Adviser Minchen Li co-authored an influential physics algorithm that Genesis uses to generate millions of high-fidelity simulated models of robot experience.
For manipulation, simulation lets teams test variations in object placement and contact conditions without repeating every experiment on physical hardware. Its value depends on how accurately those conditions are represented. Adhesives, deformable cardboard, friction, and thin materials are demanding modeling targets.
The report provides no measured simulation-to-real transfer results. The volume of simulated experience tells buyers little about how reliably a trained policy will perform at a packing station.
Architecture claims need measurements
Xian places relatively little weight on distinctions between Vision-Language-Action models and World-Action models. He argues that training infrastructure, data infrastructure, and data quality matter much more, estimating that architecture contributes perhaps 10% of the final result.
That percentage is his assessment. TechCrunch’s report supplies no controlled comparison supporting it.
His emphasis on data and infrastructure is still worth taking seriously. Observation quality, action representation, training coverage, feedback, and mechanics all affect whether a robot executes a command successfully. Model-family labels don’t capture those dependencies.
Force-rich demonstrations could be particularly useful when visual similarity conceals different physical conditions. Two boxes may look alike while requiring different grip forces. A label may appear correctly positioned but remain partly attached to its backing.
Architecture also affects what a system can represent and how it processes observations. Genesis hasn’t disclosed the model details, ablations, or task-level results needed to assess those contributions. The available evidence consists of its stated strategy and the observed demonstration.
For technical buyers, a more useful question is how performance changes when task conditions change.
Risks and limitations
Repeatability is the clearest concern. Genesis didn’t let TechCrunch rearrange the demonstration setup, and the second run exposed difficulties even without that test.
How sensitive is the policy to object placement, material variation, or an unexpected obstruction? Can it recover autonomously, and how long does recovery take? A Genesis executive described mistakes as “hiccup data,” which may help future training. Customers still have to account for failures during the current shift.
The flagship humanoid, Eno, wasn’t ready to demonstrate its capabilities to TechCrunch. Xian says the model performs better on Eno, but the observed packing task used off-the-shelf arms. Performance on that setup doesn’t establish performance on the humanoid.
Genesis says it has built 30 to 40 Enos and is piloting them with a handful of undisclosed customers. It has selected a mass-manufacturing contractor to begin work in the U.S. in 2027 and says the production design is finalized. Planned production volume hasn’t been disclosed.
Hand durability is a concrete operational problem. Xian says the dexterous hands last only a few weeks of regular work before replacement. That means recurring maintenance and expense, though the disclosed information isn’t enough to calculate either. Buyers would need replacement pricing, repair time, spare-parts availability, and a clear definition of “regular work.”
The report also provides no task success rates, cycle-time distributions, intervention frequency, or safety-validation results. Those measurements are still needed to judge deployment readiness.
What engineering teams should ask for
Teams considering this automation should evaluate complete workflows under representative conditions. Genesis’s demo links several operations, which makes it useful to assess. A failure near the end can leave a nearly finished package that still can’t ship.
The purchasing questions should be specific:
- How often does the full workflow finish without human intervention?
- What happens after a dropped box or a failed label peel?
- How do throughput and success rates change across materials and object positions?
- How much scheduled and unscheduled maintenance do the hands require?
- What safety constraints apply around workers?
Integration needs scrutiny too. The demonstration includes scanning boxes and applying a label, but the report doesn’t explain interfaces to inventory systems, job queues, or production software. A deployed packing workflow has to connect those systems as well as handle the objects.
Genesis has a credible technical approach in contact-rich data and hardware designed to use it. Sustained customer operation should provide the next evidence, including intervention counts and maintenance records. Buyers need to know both how reliably the hands work and how often they’ll need replacing.
Useful next reads and implementation paths
If this topic connects to a real workflow, these links give you the service path, a proof point, and related articles worth reading next.
Design controlled AI systems that reason over tools, environments, and operational constraints.
How field workflows improved throughput and dispatch coordination.
Mach9 has raised a $34 million Series A to build AI-powered CAD software for surveying and civil engineering. The company says nearly 100 customers have used its private beta on more than 2,500 projects, suggesting demand beyond a successful demo. Fi...
Mbodi is heading to TechCrunch Disrupt 2025 with a clear claim: industrial robots can be trained from natural-language instructions, adapt on the job, and avoid a full rework every time a packaging line changes. That matters because the bottleneck in...
Genesis AI has unveiled its first robotics model, GENE-26.5, and the bigger signal is the hardware around it. The company is building its own hand, sensing setup, and data pipeline alongside the model. Genesis emerged from stealth last year with a ...