How GENE-26.5 Is Redefining Human-Level Robot Manipulation

A robot that cracks eggs, pipettes liquids into centrifuges, solves a Rubik's Cube bimanually, and wraps wire harnesses, all using the same model weights, same hardware, same control stack.

That is not a research prototype built for a single task. That is Genesis AI's GENE-26.5, released in May 2026 as the first model in the GENE foundation model family. It is the clearest public demonstration yet that human-level dexterous manipulation is no longer a distant goal.

video source

Why Manipulation Is the Real Problem in Robotics

Navigation is mostly solved. Locomotion has checkpoints. Manipulation is still open.

Here is why. Navigation treats the world as obstacles and free space, the robot avoids contact. Locomotion uses contact for support, but the ground is stable and errors recover. Manipulation is different. Contact is the task itself.

The robot must predict interaction outcomes, reason about unknown shape, weight, and friction, and execute force and timing to the millimeter. One wrong move compounds across a long task horizon.

Over 80% of physical labor is manipulation. Almost none of it has ever been recorded as usable robot training data.

Genesis AI built GENE-26.5 around one conviction: if a robot can reliably control physical interaction with the world, everything else is support infrastructure.

The Five Axes That Define Dexterous Manipulation

GENE-26.5 is not evaluated on toy benchmarks. Genesis defines manipulation capability across five precise axes:

1. Spatial Precision - How accurately contacts and tools must be placed
2. Temporal Composition - When and how fast actions must execute to produce the right dynamics
3. Contact Richness - Number and diversity of simultaneous contacts
4. Contact Coordination - How tightly multiple contacts must synchronize
5. Tool-Mediated Interaction - Using objects as designed and in novel but physically reasonable ways

Each demo task in GENE-26.5's evaluation suite is designed to stress a different combination of these axes. None of them are isolated lab demos.

What GENE-26.5 Can Do: The Task Suite

All tasks run at 1× real-world speed using a single shared-weight model.

Cooking

A four-minute, 20+ subtask sequence in an unsimplified kitchen. The robot cracks a single egg with one hand, uses a knife and spatula, reorients a tomato for precision cutting with the opposite hand, seasons with a salt mill, and whisks. When transferring diced tomatoes, it adapts, pressing the knife flat against the cutting board as a surface and scooping with coordinated bimanual motion.

Lab Pipetting

Millimeter-level precision in a laboratory workflow. The robot picks up a pipette, inserts a tip, transfers liquid from a beaker into a tube, screws on a 1 cm cap, nudges the centrifuge "open" button, and places the tube in the rotor. It then re-grasps the pipette to hang it back on the rack. This is real lab protocol, not a simplified version.

Solving a Rubik's Cube

An external solver generates closed-loop move commands in real time. Those commands are translated to language instructions and executed by the model. Both hands work in tight coordination across every rotation. To the best of Genesis AI's knowledge, this is the first time a general-purpose bimanual robot has solved a Rubik's Cube without a specially designed mechanical fixture.

Wire Harnessing

A high-priority task in automotive manufacturing. The robot bundles deformable cables, hangs them on stands, and wraps them in tape — handling soft, highly deformable objects that have resisted robotic automation for decades.

Multi-Object Grasping

Four objects of different sizes, four distinct grasp types, one hand, one motion. The robot sorts all four into matching bins simultaneously. This demonstrates what a high-DoF hand enables beyond what standard two-finger grippers can do.

Piano Playing (Control Stack Stress Test)

This task is trained separately using reinforcement learning in simulation, guided by human demonstrations. It exists to validate the control stack's high-speed tracking accuracy, not the manipulation model. The robot plays Ferris Wheel and a clip of Rush E.

For most of these tasks, GENE-26.5 requires less than one hour of task-specific robot data, under 200 episodes for skills under 20 seconds duration.

The Full-Stack Architecture of GENE-26.5

GENE-26.5 is not a model. It is a system. Genesis treats each layer as load-bearing.

Genesis Hand 1.0

The hand is the foundation. Genesis Hand 1.0 is a direct-drive robotic hand with 20 active, back-drivable degrees of freedom. It is sized 1:1 with a human hand in dimensions and kinematic structure. The palm and fingers are covered in soft material to replicate the contact physics of human skin.

This matters because it removes the embodiment gap at the hardware level. Human hand motions map directly to the robot hand without complex retargeting algorithms. The result is near-lossless transfer from human demonstration to robot execution.

The Human-Centric Data Engine

Genesis uses three complementary data sources:

• Glove data : EMF-based finger tracking with dense tactile sensing. Minimally invasive. Integrates into real workflows so real work becomes data collection with zero overhead.
• Egocentric video : Captures natural human behavior and real-world task diversity at scale
• Third-person video : Internet-scale coverage of physical interaction

Together, Genesis has collected over 200,000 hours of data across these modalities in collaboration with partners.

  robot data sources

The glove interface is shared between human and robotic hand, meaning the same device captures data and then maps directly to deployment. Fidelity is preserved end-to-end.

The Foundation Model

GENE is designed around a joint distribution over trajectories. It uses flow matching to capture multimodal futures while preserving temporal dynamics. It trains across five modalities: language, vision, proprioception, tactile, and action, without requiring explicit alignment between them.

The model handles multiple query types from this joint distribution: control, goal inference, state estimation, inverse dynamics, and value estimation. Missing modalities are inferred through denoising.

It imports scale from two external sources:

• Vision-Language Models (VLMs) for semantic intent and representation
• World Models (action-conditioned video generation) for physical and temporal dynamics

Control Stack: 3 ms End-to-End Latency

Standard robotic arms ship with vendor controllers that introduce latency and tracking error. When training from human motion rather than robot teleoperation, that mismatch becomes a hard problem. The model learns from data that does not reflect the robot's actual dynamics at deployment.

Genesis replaced the vendor controller entirely. Their custom control middleware achieves:

• End-to-end latency as low as 3 ms under tuned settings
• 500 Hz control loop on both arms through a single EtherCAT Y-slave network
• PREEMPT_RT kernel with isolated CPU cores for deterministic real-time execution
• KickCAT as the EtherCAT master with Distributed Clocks support
• Both position and impedance control with position and velocity targets

The difference is measurable. Tracking a 15 cm circle over 4 seconds, the default vendor controller produces an average error of ~20 mm. Genesis's controller reduces that to ~2 mm, an order-of-magnitude improvement. On single-joint sinusoidal tracking, default latency is ~80 ms. Genesis achieves 9 ms, tunable down to 3 ms.

This is what closes the human-to-robot gap at the source.

Scaling Laws Apply to Robotics Too

Genesis validates GENE-26.5's scaling behavior in three stages.

Open-loop pre-training: Increasing model size and compute consistently reduces validation loss. Larger models reach lower asymptotic error. This confirms that standard scaling laws from language model research apply to robotic foundation models.

  Scaling curve plot (model size vs validation loss)

Closed-loop simulation evaluation: Genesis uses Genesis World for simulation-based closed-loop evaluation, with zero simulation training data. The realism level of Genesis World is high enough to evaluate models trained only on real-world data. Each data point in their scaling plot represents 200 evaluation setups and over 150 robot-hours. The full plot would require 2,700 human-robot hours to run in the real world. Simulation makes this tractable.

The key finding: scaling pre-training data leads to stronger zero-shot generalization.

Real-world fine-tuning: Tasks excluded from pre-training are evaluated with ~20-30 minutes of task-specific data. More pre-training data means faster adaptation, less data needed, and higher final task success rate.

  Real-world fine-tuning success rate plot from

The trend is consistent across all three stages: scale improves both generalization and adaptation efficiency.

Conclusion

GENE-26.5 is an early release. Genesis AI calls it a beginning. But it establishes every pillar needed to scale: biomimetic hardware, high-fidelity human data, a low-latency custom control stack, a multimodal foundation model, and simulation infrastructure for reproducible evaluation at scale.

The path to general-purpose robots runs through manipulation. GENE-26.5 is the clearest evidence yet that the full-stack approach, not pure model training, is what moves that path forward.

Power Your Robot AI Pipeline with Labellerr

Training a manipulation foundation model like GENE-26.5 starts with data. Specifically, high-quality egocentric and multimodal data that captures real human behavior.

Labellerr specializes in exactly this:

• Egocentric data annotation - Frame-level labeling of first-person video, hand pose, and interaction sequences for robotic pre-training
• Multimodal data pipelines - Coordinated annotation across video, tactile, proprioceptive, and language streams
• High-throughput annotation at scale - Built for foundation model data volumes, not one-off datasets

Whether you are collecting glove demonstrations, annotating industrial task videos, or building evaluation sets for sim-to-real transfer, Labellerr gives your team the infrastructure to move fast without compromising label quality.

Talk to the Labellerr team about your robotics data pipeline →

FAQs