Sim-to-Real Transfer and Domain Randomization

Because the machines are not yet built, most policies are born in simulation; domain randomization and calibrated twins narrow the gap so they survive first contact with hardware.

Learning before there is a machine

Breeder construction begins Q2 2027 and first tritium is targeted near 2030, so today every control policy is trained against simulation and the digital twin, not hardware. The central risk is the reality gap: a policy that is optimal in simulation can fail on a real machine whose behavior differs in ways the simulator got subtly wrong. Sim-to-real MLOps exists to shrink and characterize that gap.

Domain randomization is the primary tool: rather than training against one nominal simulator, policies train across a distribution of simulators with randomized physics parameters, sensor noise, latencies, and actuator responses. A policy that performs across the whole distribution is far likelier to perform on the one real instance, which is treated as another draw from the distribution.

Techniques

• Domain randomization over uncertain physics and hardware parameters
• Twin calibration to real diagnostics as data arrives (breeder from ~2030)
• System-identification to tighten the simulator toward the true machine
• Robustness margins: require performance at distribution tails, not just the mean
• Explicit reality-gap metrics tracked over commissioning

def randomized_env():
    return Sim(
      tau_E   = sample(0.9, 1.1) * TAU_E_NOMINAL,   # confinement
      delay   = sample(0, 3) * CYCLE,               # actuation lag
      noise   = sample_noise_model(),
      # burner: plug field spread bounded by 26.49 T design point
      # breeder: triangularity jitter around delta -0.30
    )
# train policy across many draws -> robust to the real one

Sim-to-real is inseparable from twin refinement: the same measured-vs-predicted residuals that refine the twin also tighten the simulator distribution. For the burner, honest limits apply — the plug regime is 166 to 830x beyond any device, so there is no near-term hardware to transfer to, and burner policies remain simulation studies validated only within provable bounds.