Genesis AI Shows Embodied Model That Cooks, Experiments and Plays Piano

Genesis AI, after open‑sourcing its Genesis physics engine, has unveiled its first embodied foundation model, GENE‑26.5. The model can autonomously perform long‑horizon tasks such as cooking a meal, solving a Rubik’s cube, playing the Rush E piano piece, laboratory pipetting and cable‑bundle organization.

Task details : The cooking demo lasts about four minutes and comprises more than twenty sub‑tasks, requiring single‑hand egg beating, tomato cutting, use of a towel, salt grinder, whisk, knife, spatula and pan, as well as coordinated two‑hand actions. The lab‑pipetting scenario demands millimetre‑level precision, tool use, and stable handling of small objects. The Rubik’s‑cube solution is achieved without special grippers, using an external solver to generate motion commands. The piano performance showcases speed beyond typical human limits.

Data efficiency : Most of these challenging skills need less than one hour of robot‑specific data—fewer than 200 episodes for skills under twenty seconds—demonstrating that the model can be aligned with a small amount of task‑specific data after pre‑training on human demonstration data.

Full‑stack approach : Genesis argues that a model alone is insufficient. Their system integrates a human‑size hand (Genesis Hand 1.0 with 20 active DOF and flexible skin), a low‑cost human data‑collection pipeline, a custom simulation‑evaluation stack, a multimodal foundation model, and a low‑latency, high‑precision control system. This “full‑stack robotics” is intended to narrow the embodiment gap.

Simulation as accelerator : Real‑world evaluation is slow; a single robot and human evaluator can only run one task at a time. Genesis therefore built a closed‑loop simulation core that can vary lighting, background, object properties, scene configuration and task instructions. Each data point in their benchmark represents 200 simulation settings and over 150 hours of robot execution; reproducing the same in the real world would require roughly 2 700 hours of human‑robot testing.

Human‑centric data engine : Three data streams are collected: (1) glove data capturing high‑precision hand motion and tactile signals; (2) first‑person video of natural task execution; (3) third‑person video from internet‑scale sources covering diverse physical interactions. Their self‑developed data glove uses EMF finger tracking and dense tactile sensing, allowing unobtrusive capture of skilled human work.

Control stack redesign : To eliminate latency, tracking error and controller artifacts, Genesis replaced the off‑the‑shelf robot controller with a custom middleware. The new system achieves end‑to‑end latency as low as 3 ms and reduces the average error on a 15 cm circular‑trajectory test from about 20 mm (stock controller) to roughly 2 mm, an order‑of‑magnitude improvement.

Roadmap : The “26.5” in GENE‑26.5 points to May 2026. Genesis states this is only the first version; future iterations will include a full‑body robot and rapid model updates, with simulation‑driven evaluation remaining the primary bottleneck. Investors have called the demonstration “the strongest general‑operation demo”.

Reference: https://www.genesis.ai/blog/gene-26-5-advancing-robotic-manipulation-to-human-level