🌉 Sim-to-Real: Domain Randomization and Transfer

O reality gap is the systematic divergence between simulator behavior and the physical world's behavior. A policy that learns to exploit simulator imperfections—idealized friction, perfect contact, noise-free sensors—performs brilliantly in simulation and fails on hardware. Mapping the four sources understanding the gap is the first step toward closing it.

A domain randomization (DR) reverses the problem: instead of making the sim perfect, it makes it variable enough that the real world looks like just another sample from the training distribution. If the policy has seen 10,000 variations in friction, light, and mass, real-world friction falls within what it already handles. It’s the most robust and affordable technique for transfer zero-shot.

# Pseudo-DR por episódio (Isaac Lab / MuJoCo)
def randomize(env):
    env.friction   = U(0.4, 1.2)      # atrito
    env.mass_scale = U(0.8, 1.2)      # massa do objeto
    env.light_dir  = sample_sphere()  # direção de luz
    env.texture    = random_texture() # aparência
    env.act_delay  = randint(0, 3)    # atraso de atuador (steps)
    env.obs_noise  = N(0, 0.01)       # ruído de sensor
    return env  # o real vira "mais uma amostra" disso

When the visual gap is too large to cover with noise alone, the domain adaptation comes in: instead of making the policy invariant by brute force, actively brings the distribution closer to features of sim and real. Adversarial techniques align embeddings; models such as RCAN translate real images into a sim "canonical" form before inferring.

DR covers uncertainty; system identification (sysID) reduces it at the source. By measuring the real robot—masses, centers of inertia, joint friction, control latency—you calibrate the simulator to reproduce the specific hardware. The result is a digital twin: a faithful digital twin that serves as a training ground and a safe test bench.

# sysID: ajustar parâmetros do sim ao real por otimização
real_traj = record(robot, commands)          # dados do hardware
def loss(theta):
    sim_traj = simulate(commands, params=theta)
    return mse(sim_traj, real_traj)           # quão próximo o twin está
theta_star = minimize(loss, init=spec_sheet) # calibra massa/atrito/delay
# theta_star = simulador que "vira" o robô real

The dominant practical recipe in 2026 isn't "sim OR real"—it's co-training. The sim provides coverage inexpensive (millions of variations), and a handful of real demonstrations provides fidelity final. Mixing ~95% sim + ~5% real, or doing a short fine-tuning run on real data, usually outperforms either one alone.

The most common mistake in sim-to-real is confusing sim-score with real capability. A policy that reaches 99% in simulation says little about the hardware. The only metric that matters for deployment is success rate on the physical robot, measured under variation the policy hasn't seen, using honest protocols.