Genesis AI and GENE-26.5: Toward Robots with Human-Like Movements

Genesis AI, a company founded between Paris and California, aims to revolutionize robotics by leveraging advancements in generative artificial intelligence. Théophile Gervet, co-founder of the company, emphasizes that generative AI has transformed software by enabling models to generalize cognitive tasks from massive volumes of data. Genesis AI seeks to apply this transformation to the physical world with its GENE-26.5 model, designed to equip robots with physical manipulation capabilities akin to those of humans.

The GENE-26.5 model aims to create a physical foundation model capable of learning complex manual tasks, transferring them across different environments, and generalizing motor skills at scale. However, the main obstacle to this advancement lies in the collection of physical data. Unlike language models, which benefit from vast online data corpora, robotic systems have limited exploitable physical data at scale. Human gestures, tactile interactions, object manipulation, and coordinated movements are challenging to capture, standardize, and reproduce.

Genesis AI claims to have developed an architecture that overcomes this limitation. The company has created a robotic hand that replicates human morphology, as well as a glove equipped with tactile sensors that establish a direct correspondence between the human hand, the glove, and the robotic hand. This structure aims to transform human gestures into directly usable training data for AI models. This system could significantly reduce the morphological gap that has historically limited robotic learning from human data. Genesis AI also indicates that its data collection glove would cost one hundred times less than existing solutions while significantly improving the quality and speed of data collection.

The adopted approach relies on a classic logic in artificial intelligence: accumulating massive volumes of proprietary data to train models capable of generalizing their skills. Genesis AI plans to deploy its collection devices directly at industrial partners to continuously gather data from real tasks performed by human operators. Simultaneously, the company utilizes first-person viewpoint videos and video content from the internet to feed its foundation models.

GENE-26.5 has been showcased through a series of demonstrations highlighting particularly complex tasks for robotic systems, such as meal preparation, coordinated bimanual manipulation, laboratory experiments, solving a Rubik’s Cube, and musical interpretation on the piano. These demonstrations specifically target areas where robotics still faces limitations: fine manipulation, continuous coordination, force control, dynamic adaptation, and interactions in semi-structured environments.

Genesis AI's project is part of a broader evolution in the sector. Following language models and multimodal systems, several players are attempting to develop Physical Foundation Models, which are models capable not only of understanding the world but also of physically acting within it, similar to AMI LABS. This transition could profoundly alter the industrial balance of robotics. Until now, most industrial robots relied on specialized systems designed for specific tasks in controlled environments. In contrast, robotic foundation models seek to produce more generalist systems capable of quickly adapting to new uses.

To achieve this, Genesis AI adopts a full-stack strategy that integrates hardware, AI models, data collection systems, and simulation. The company is notably developing a simulation platform aimed at reducing the “sim-to-real gap,” which is the discrepancy between performance observed in virtual environments and that achieved in the real world. Through more realistic physical and visual engines, Genesis AI aims to accelerate the training and evaluation cycles of its models without relying solely on costly physical testing.

Genesis AI plans to unveil its first general-purpose robot based on the technologies presented with GENE-26.5 soon. Behind this announcement lies an increasingly intense industrial competition surrounding the next generation of artificial intelligences capable of directly interacting with the physical world.

Co-founded by Zhou Xian and Théophile Gervet, Genesis AI raised $105 million in seed funding last year, notably from Eclipse, Khosla Ventures, Bpifrance, and HSG, with support from investors such as Eric Schmidt and Xavier Niel. This level of funding places Genesis AI among the most heavily capitalized projects in the robotics sector at their initial phase.