French robotics-focused start-up Genesis AI has released its first foundational model, GENE-26.5, along with a human-like hand built in-house that’s capable of carrying out complex tasks such as cooking or solving a Rubik’s Cube.
The company, which completed a $105 million (£77.2m) seed funding round in July of last year, began with a focus on a model to power real-world robotics, then built its own hand in order to have more accurate control over the full stack, said co-founder and chief executive Zhou Xian.
Unlike robotic hands built by dozens of other companies, which tend to have two or three fingers, the unit from Genesis AI closely resembles a human appendage, allowing data to be more directly transferred to the device. This design choice is critical for bridging the gap between simulated training and real-world manipulation. Most existing robotic hands are either too simplistic (e.g., parallel-jaw grippers) or too complex (e.g., multi-fingered hands that are difficult to control). Genesis AI’s approach aims for a middle ground – a hand that is anatomically plausible yet robust enough for industrial use.
The GENE-26.5 model itself is a large foundational model trained on diverse robotic and human demonstration data. Its number ‘26.5’ likely refers to the parameter count in billions, though the company has not disclosed exact details. Foundational models in robotics are analogous to large language models in AI – they serve as a base that can be fine-tuned for specific tasks. Genesis AI’s model is designed to generalize across different hardware platforms, but the company decided to build its own hand to validate the model’s capabilities and optimize the entire hardware-software stack. This vertical integration is reminiscent of how Tesla builds its own chips or how Apple controls both hardware and software.
The hand’s human-like dexterity opens up applications that require fine motor skills. In a demonstration video, Genesis showed a pair of its hands carrying out complex or delicate processes such as preparing a smoothie or playing the piano. The hand can also manipulate small objects like a Rubik’s Cube, rotating and aligning it with precision. Such tasks were previously only possible with specialized robotic arms in controlled environments.
Genesis has also created lightweight sensor gloves that can be worn by workers in fields such as pharmaceuticals or manufacturing, capturing data that can be passed along to the AI model. These gloves are equipped with touch sensors that record finger positions, forces, and grip patterns. This data collection method is less intrusive than traditional motion capture systems and allows the company to gather large amounts of human demonstration data quickly. However, whether workers would be happy to help train a potential robotic replacement remains in question. This raises ethical considerations about the future of human labor and the role of such data collection.
Data from the gloves can be combined with information gleaned from videos of workers carrying out tasks, Genesis said. This multimodal learning approach – combining tactile and visual data – is at the cutting edge of robotics research. By training on both, the AI model learns not just the visual sequence of actions but also the haptic feedback required for delicate manipulations, such as picking up an egg without breaking it or handling a tiny screw.
Co-founder and president Theophile Gervet previously worked for France’s Mistral, a leading AI startup in Europe. This connection highlights the cross-pollination between large language model expertise and robotics. Many researchers from the natural language processing and computer vision fields are now applying their knowledge to robotics, seeing it as the next frontier for AI. The company retains a substantial presence in Europe, which it attributed to the continent’s large talent base and potential market of industrial customers. Europe has a strong manufacturing sector, especially in Germany, France, and Italy, and many factories are looking to automate repetitive or dangerous tasks.
Genesis AI said its team of 60 people is split roughly evenly between the US and Europe, with a slightly larger group in the US base in California. In addition to San Carlos, California, it has offices in Paris and London. This geographic distribution allows the company to tap into the venture capital ecosystem of Silicon Valley while maintaining close contact with European industrial clients.
The company said it is working to supplement the mechanical hand with a full-body, general purpose robot. This robot would likely take the form of a humanoid, capable of walking, reaching, and handling tasks that require full-body coordination. Humanoid robotics have seen a resurgence in recent years, with companies like Boston Dynamics, Tesla (Optimus), and Figure AI making headlines. Genesis AI’s focus on dexterous manipulation could give it an edge in tasks that require fine motor control, such as assembly line work, surgery assistance, or household chores.
The seed funding round last year was co-led by Eclipse and Khosla Ventures, with additional backing from Bpifrance, HSG, former Google chief Eric Schmidt, telecoms entrepreneur Xavier Niel, MIT Computer Science and Artificial Intelligence Laboratory director Daniela Rus and Apple distinguished scientist Vladlen Koltun. This star-studded investor list reflects the high expectations for the field of general-purpose robotics. Eric Schmidt and Xavier Niel bring strategic advice, while Daniela Rus and Vladlen Koltun provide technical credibility.
The development of foundational models for robotics echoes the trajectory of natural language processing. Just as GPT-3 and its successors have enabled a wide range of language tasks, models like GENE-26.5 could unlock a new generation of robots that can follow natural language commands and adapt to new situations without explicit programming. However, challenges remain. Real-world environments are messy and unpredictable; sensors can be noisy; and safety is paramount when robots operate near humans.
One of the key technical hurdles is sim-to-real transfer – training a model in simulation and having it work in the physical world. Genesis AI likely uses large-scale simulations to generate training data, then fine-tunes with real-world data from the gloves and videos. The human-like hand design helps because it reduces the discrepancy between human data (collected via gloves) and robot execution. If the robot hand had a different number of fingers or range of motion, transferring human skills would require complex retargeting algorithms.
The launch of GENE-26.5 comes at a time when several other startups are also pushing the boundaries of dexterous manipulation. Companies like Shadow Robot (UK), Robotiq (Canada), and Franka Emika (Germany) offer robot hands, but few have attempted to build a full-stack AI model from scratch. By controlling both the model and the hardware, Genesis AI can optimize performance end-to-end, potentially achieving faster development cycles and better integration.
The market for robotic hands is projected to grow significantly in the coming years, driven by demand from logistics, manufacturing, and healthcare. According to a report by MarketsandMarkets, the global robotic gripper market is expected to reach $10.5 billion by 2027, up from $5.3 billion in 2022. However, most of that market is for simple grippers; dexterous hands are still a niche. If Genesis AI can demonstrate reliable performance in real-world applications, it could capture a significant share of the premium end of the market.
Another important aspect is the cost of the robot hand. Building a highly dexterous hand with many sensors and actuators is expensive. The Genesis AI hand is likely to be priced in the tens of thousands of dollars, which may limit its adoption to high-value applications like research labs, pharmaceutical manufacturing, or specialized assembly tasks. However, as production scales and costs come down, broader adoption could follow.
The company also faces competition from big tech firms. Google, Amazon, and Microsoft have all invested in robotics research. Google’s RT-2 model (Robotic Transformer 2) and Amazon’s work on warehouse robotics are notable. However, these projects are often internal and not commercialized as standalone products. Genesis AI’s startup agility could allow it to move faster and target specific verticals.
In terms of regulatory environment, the European Union is developing the AI Act which may have implications for robotics, especially regarding safety and transparency. Genesis AI’s European presence positions it to navigate these regulations. Additionally, data privacy laws like GDPR will affect the collection of data from workers wearing sensor gloves. The company will need to ensure that data is anonymized and used ethically.
The future of Genesis AI looks promising but uncertain. The company has strong backing, a clear technical vision, and a first product that demonstrates impressive capability. However, transforming a demonstration into a reliable, cost-effective product for industrial customers is a different challenge. The team’s experience from Mistral and other top AI labs gives it a good chance. Moreover, the trend towards humanoid robots and general-purpose AI suggests that the market is ready for such innovations.
As Genesis AI works on its full-body robot, it will likely leverage the same foundational model, GENE-26.5, as the “brain” that controls both the hands and the rest of the body. This unified control architecture could simplify development and enable more complex behaviors. In the long term, the goal is to have a robot that can learn new tasks simply by watching humans, much like a human apprentice learns by observing a master craftsman. The sensor glove data provides a direct channel for such imitation learning.
The success of Genesis AI could also spur more investment into the robotics AI stack. Currently, many robotics companies treat hardware and software as separate entities, leading to integration issues. If Genesis proves that a tightly integrated approach works, others may follow. This could accelerate the timeline for commercially viable general-purpose robots. The race is on, and Genesis AI has just made a bold move.
Source: Silicon UK News