Cyclotron

Cyclotron is the locomotion + manipulation training pipeline—a core differentiator that improves reliability and reduces data requirements for humanoid robot training.

The Problem We’re Solving

Training humanoid robots in the real world is slow, expensive, and dangerous. Real-world data collection requires physical robots, which are costly to build and maintain. Failures during training can damage hardware and delay timelines. The traditional approach to robotics development—train, break, repair, repeat—is unsustainable at scale.

The Cyclotron Solution

Cyclotron uses sim-to-real transfer via domain randomization to reduce real-world training needs. This is a well-established approach that acknowledges the reality gap while making failure cheap and iteration fast.

What Cyclotron Provides

• Digital twin infrastructure: High-fidelity virtual replicas of humanoid platforms for testing and training
• Physics modeling: Accurate simulation of the physical dynamics of Asimov, including mass distributions, joint stiffness, friction, and contact dynamics
• Electronics noise simulation: Models the sensor and actuator noise from Asimov’s electronic systems, narrowing the sim-to-real gap
• Domain randomization: Varying physical and electronic parameters during training to make policies robust to reality gaps
• Locomotion pipeline: Training policies for stable, efficient walking across varied terrain
• Manipulation pipeline: Training policies for grasping, handling, and manipulating objects
• Whole-body coordination: Training policies that integrate locomotion and manipulation simultaneously

Narrowing the Sim-to-Real Gap

Cyclotron goes beyond generic simulation by modeling the specific physics and electronics characteristics of Asimov. This includes:

• Actuator dynamics: Motor behavior, torque curves, and response latencies
• Sensor characteristics: Noise profiles, calibration offsets, and sampling rates
• Material properties: Friction coefficients, compliance, and contact dynamics

By incorporating these details, policies trained in Cyclotron transfer more reliably to physical Asimov robots, reducing the real-world training needed to achieve reliable performance.

Training Outputs

Trained policies from Cyclotron deploy directly to physical robots via Agent Platform. The output is a verified policy ready for real-world testing, with confidence metrics based on simulation performance.

Why Cyclotron Matters

Cost reduction: Simulation training reduces real-world training needs by orders of magnitude.

Reliability: Domain randomization and accurate physics modeling produce policies robust to real-world variation.

Faster training: Electronics noise and physics fidelity mean policies converge faster with fewer sim-to-real iterations.

Safety: Simulation allows testing dangerous scenarios without risk of hardware damage.

Iteration speed: Policy changes are validated in hours rather than days or weeks.

Integration

Cyclotron connects:

• Agent Platform: Receives policies for simulation validation
• Uranus: Receives simulation data for curriculum building and comparison with real-world performance
• Asimov: Outputs trained policies for physical deployment