Trained in simulation, against real constraints
Simulation lets a policy fail as often as it needs to at no cost. It is only useful when the environment encodes what the real machine can and cannot do. Our training environments are built from actual machine specifications and kept honest through the calibration described on our simulation and real-to-sim page. Where the dynamics are too complex to write down, learned robotic world models stand in for hand-built physics.
The constraints we model include:
- Actuator limits: rate, force, and range bounds taken from the machine, not tuned for convenience
- Sequencing rules: operations the hardware requires in a fixed order
- Load-dependent dynamics: behavior that shifts with payload and configuration
- Command rates: the update frequency the real control path supports
Deterministic gates before hardware
Every trained policy is verified by running it, never by reading a training curve. A gate is a scripted evaluation whose verdict is computed in code: the policy executes a fixed set of scenarios, and pass or fail comes from measurable criteria such as tracking error, limit violations, and task completion. Each gate is confirmed to fail a bad policy before it is trusted to pass a good one, and it is never weakened to let a marginal policy through.
Only a policy that goes green moves toward hardware, and it is deployed under the same deterministic gates that govern every GENOMA model. This is the same verify-by-running discipline GENOMA applies to every model it ships.
From data to deployed policy
Policies need task distributions and demonstrations, not just physics. Craneformerz, our AI videogame and data-acquisition platform for robotic lifting, turns human play into training and evaluation data for lifting tasks. Reinforcement learning then owns the layer where precision matters most: low-level control that respects every limit listed above.
Above that layer sit vision-language-action models, which turn perception and instructions into task-level commands for the control policies to execute. Training and serving both run on hardware GENOMA owns, so the full loop from simulation to deployed policy stays under one roof.
Questions
Why train heavy-machinery policies in simulation instead of on the real machine?
Real machines are expensive to run, slow to reset, and dangerous to explore on. In simulation a policy can fail thousands of times at no cost, with the machine's actual limits encoded so it learns them from the start. The gap to reality is managed through real-to-sim calibration and deterministic checks before any hardware run.
What is a deterministic gate?
A scripted evaluation whose verdict is computed in code, not judged by a person or another model. The policy runs a fixed set of scenarios and passes or fails on measurable criteria. A policy that fails does not ship, and the gate is never weakened to make it pass.
How does reinforcement learning fit with vision-language-action models?
They are complementary layers. Reinforcement learning produces the low-level control policies that respect actuator and sequencing limits, while vision-language-action models translate perception and instructions into the task-level commands those policies execute. Together they cover the stack from intent to actuation.