Entropy-Driven Sim2Real Transfer
Trit: -1 (MINUS - analysis/verification) Color: #E85B8E (Rose Pink) URI: skill://entropy-sim2real#E85B8E
Core Insight
Entropy bridges the sim-real gap by:
- Maximizing entropy in simulation → Policy sees diverse conditions
- Minimizing entropy at deployment → Uncertainty collapses to reality
- Information-theoretic alignment → Match distributions, not parameters
SIMULATION REALITY
High Entropy ─────────────────────────────▶ Low Entropy
H(params) = max ══════════▶ H(params) ≈ 0
H(π|s) = high ══════════▶ H(π|s) = focused
p(sim) = broad ══════════▶ p(real) = delta
┌─────────────────┐ ┌─────────────────┐
│ MANY POSSIBLE │ BRIDGE │ ONE ACTUAL │
│ WORLDS │───────────────│ WORLD │
│ (superpos.) │ │ (collapsed) │
└─────────────────┘ └─────────────────┘
Three Entropy Mechanisms
1. Domain Randomization Entropy
Maximize entropy over simulation parameters:
import jax
import jax.numpy as jnp
from typing import Dict
class EntropyMaximizingRandomizer:
"""Domain randomization that maximizes parameter entropy."""
def __init__(self, param_ranges: Dict[str, tuple]):
self.param_ranges = param_ranges
def entropy(self, distribution: str = "uniform") -> float:
"""Compute entropy of parameter distributions."""
H = 0.0
for name, (low, high) in self.param_ranges.items():
if distribution == "uniform":
# H(Uniform) = log(b - a)
H += jnp.log(high - low)
elif distribution == "gaussian":
# H(Gaussian) = 0.5 * log(2πeσ²)
sigma = (high - low) / 4 # 95% within range
H += 0.5 * jnp.log(2 * jnp.pi * jnp.e * sigma**2)
return H
def sample(self, key: jax.random.PRNGKey) -> Dict[str, float]:
"""Sample parameters to maximize coverage."""
params = {}
for i, (name, (low, high)) in enumerate(self.param_ranges.items()):
k = jax.random.fold_in(key, i)
# Uniform maximizes entropy for bounded support
params[name] = jax.random.uniform(k, minval=low, maxval=high)
return params
def adaptive_entropy(
self,
key: jax.random.PRNGKey,
real_samples: jnp.ndarray,
temperature: float = 1.0
) -> Dict[str, float]:
"""
Adapt randomization to maximize coverage of real distribution.
Uses maximum entropy principle: find distribution with highest
entropy subject to matching observed moments.
"""
# Estimate real distribution moments
real_mean = jnp.mean(real_samples, axis=0)
real_var = jnp.var(real_samples, axis=0)
# Maximum entropy distribution matching moments = Gaussian
params = {}
for i, (name, _) in enumerate(self.param_ranges.items()):
k = jax.random.fold_in(key, i)
# Sample from Gaussian matching real moments (max entropy)
params[name] = jax.random.normal(k) * jnp.sqrt(real_var[i]) + real_mean[i]
return params
2. Maximum Entropy RL
Policy optimization with entropy regularization:
class MaxEntropyPPO:
"""
PPO with entropy bonus for robust sim2real.
Objective: max E[Σ γᵗ(rₜ + α·H(π(·|sₜ)))]
High entropy → diverse actions → robust to perturbations
"""
def __init__(
self,
entropy_coef: float = 0.01,
target_entropy: float = -1.0,
auto_tune: bool = True
):
self.alpha = entropy_coef
self.target_entropy = target_entropy
self.auto_tune = auto_tune
if auto_tune:
# Learnable temperature (SAC-style)
self.log_alpha = jnp.log(entropy_coef)
def policy_entropy(self, logits: jnp.ndarray) -> float:
"""Compute policy entropy H(π) = -Σ π(a)log(π(a))."""
probs = jax.nn.softmax(logits)
log_probs = jax.nn.log_softmax(logits)
return -jnp.sum(probs * log_probs, axis=-1).mean()
def gaussian_entropy(self, std: jnp.ndarray) -> float:
"""Entropy of Gaussian policy: H = 0.5 * log(2πeσ²)."""
return 0.5 * jnp.log(2 * jnp.pi * jnp.e * std**2).sum(axis=-1).mean()
def entropy_loss(
self,
policy_entropy: float,
update_alpha: bool = True
) -> tuple:
"""
Compute entropy loss and optionally update temperature.
We want: H(π) ≥ H_target
Loss: α * (H(π) - H_target)
"""
entropy_bonus = self.alpha * policy_entropy
if self.auto_tune and update_alpha:
# Dual gradient descent on temperature
alpha_loss = -self.log_alpha * (policy_entropy - self.target_entropy)
return entropy_bonus, alpha_loss
return entropy_bonus, 0.0
def robust_policy_loss(
self,
advantages: jnp.ndarray,
log_probs: jnp.ndarray,
old_log_probs: jnp.ndarray,
policy_entropy: float,
clip_ratio: float = 0.2
) -> float:
"""
PPO loss with entropy regularization.
L = L_clip + α·H(π)
High entropy prevents overconfident policies that
fail on real hardware.
"""
# Standard PPO clipped objective
ratio = jnp.exp(log_probs - old_log_probs)
clipped = jnp.clip(ratio, 1 - clip_ratio, 1 + clip_ratio)
policy_loss = -jnp.minimum(ratio * advantages, clipped * advantages).mean()
# Entropy bonus (negative because we minimize loss)
entropy_bonus = -self.alpha * policy_entropy
return policy_loss + entropy_bonus
3. Information-Theoretic Bridging
Minimize information gap between sim and real:
class InformationTheoreticBridge:
"""
Bridge sim and real via information-theoretic measures.
Key insight: We can't match physics exactly, but we can
match the *information content* of observations.
"""
def mutual_information(
self,
sim_obs: jnp.ndarray,
real_obs: jnp.ndarray
) -> float:
"""
Estimate I(sim; real) - how much sim tells us about real.
High MI = sim is predictive of real (good!)
Low MI = sim and real are independent (bad!)
"""
# Use MINE estimator or simple correlation
joint_cov = jnp.cov(sim_obs.T, real_obs.T)
n = sim_obs.shape[1]
cov_sim = joint_cov[:n, :n]
cov_real = joint_cov[n:, n:]
cov_joint = joint_cov
# MI = 0.5 * log(|Σ_sim||Σ_real| / |Σ_joint|)
mi = 0.5 * (
jnp.linalg.slogdet(cov_sim)[1] +
jnp.linalg.slogdet(cov_real)[1] -
jnp.linalg.slogdet(cov_joint)[1]
)
return mi
def domain_divergence(
self,
sim_obs: jnp.ndarray,
real_obs: jnp.ndarray,
method: str = "wasserstein"
) -> float:
"""
Measure divergence between sim and real distributions.
Lower divergence = better sim2real transfer.
"""
if method == "kl":
# KL(real || sim) - how surprised is sim by real?
# Requires density estimation
pass
elif method == "wasserstein":
# W_2 distance (optimal transport)
mu_sim = jnp.mean(sim_obs, axis=0)
mu_real = jnp.mean(real_obs, axis=0)
cov_sim = jnp.cov(sim_obs.T)
cov_real = jnp.cov(real_obs.T)
# W_2² = ||μ_sim - μ_real||² + Tr(Σ_sim + Σ_real - 2(Σ_sim^½ Σ_real Σ_sim^½)^½)
mean_diff = jnp.sum((mu_sim - mu_real)**2)
# Simplified: use Frobenius norm of covariance difference
cov_diff = jnp.sum((cov_sim - cov_real)**2)
return jnp.sqrt(mean_diff + cov_diff)
elif method == "mmd":
# Maximum Mean Discrepancy
from functools import partial
def rbf_kernel(x, y, sigma=1.0):
return jnp.exp(-jnp.sum((x - y)**2) / (2 * sigma**2))
n, m = len(sim_obs), len(real_obs)
# MMD² = E[k(x,x')] + E[k(y,y')] - 2E[k(x,y)]
xx = jnp.mean(jax.vmap(lambda x: jax.vmap(lambda x2: rbf_kernel(x, x2))(sim_obs))(sim_obs))
yy = jnp.mean(jax.vmap(lambda y: jax.vmap(lambda y2: rbf_kernel(y, y2))(real_obs))(real_obs))
xy = jnp.mean(jax.vmap(lambda x: jax.vmap(lambda y: rbf_kernel(x, y))(real_obs))(sim_obs))
return xx + yy - 2 * xy
def entropy_matching_loss(
self,
sim_obs: jnp.ndarray,
real_obs: jnp.ndarray
) -> float:
"""
Match entropy profiles between sim and real.
If H(sim) >> H(real): sim too noisy, reduce randomization
If H(sim) << H(real): sim too deterministic, increase randomization
"""
def estimate_entropy(obs):
# Estimate via covariance determinant (Gaussian assumption)
cov = jnp.cov(obs.T)
return 0.5 * jnp.linalg.slogdet(cov)[1]
H_sim = estimate_entropy(sim_obs)
H_real = estimate_entropy(real_obs)
return (H_sim - H_real)**2
The Entropy Bridge Pipeline
┌────────────────────────────────────────────────────────────────────┐
│ ENTROPY-DRIVEN SIM2REAL │
├────────────────────────────────────────────────────────────────────┤
│ │
│ PHASE 1: Maximum Entropy Simulation │
│ ──────────────────────────────────── │
│ │
│ Domain Params Policy Observations │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ H(θ) = max │ ───▶ │ H(π|s) = αT │ ───▶ │ H(o) = high │ │
│ │ friction ∈ │ │ explore all │ │ diverse │ │
│ │ [0.3, 1.5] │ │ actions │ │ experiences │ │
│ │ mass ∈ │ └─────────────┘ └─────────────┘ │
│ │ [0.8, 1.2] │ │
│ └─────────────┘ │
│ │
│ PHASE 2: Information Bridge │
│ ─────────────────────────── │
│ │
│ Sim Distribution Divergence Real Distribution │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ p(o|sim) │ ──────▶│ W(sim,real) │◀─── │ p(o|real) │ │
│ │ (broad) │ │ minimize │ │ (narrow) │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │ │
│ Adapt randomization │
│ to match real entropy │
│ │
│ PHASE 3: Entropy Collapse at Deployment │
│ ──────────────────────────────────────── │
│ │
│ Policy trained on Deployed on Result │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ ALL possible│ ───▶ │ ONE actual │ ───▶ │ ROBUST to │ │
│ │ worlds │ │ world │ │ any world │ │
│ │ (superpos.) │ │ (collapsed) │ │ in support │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │
└────────────────────────────────────────────────────────────────────┘
Integration with K-Scale Stack
from ksim import PPOTask, PhysicsRandomizer
from ksim.randomizers import (
StaticFrictionRandomizer,
MassMultiplicationRandomizer,
JointDampingRandomizer,
)
class EntropyBridgedKBotTask(PPOTask):
"""K-Bot training with entropy-driven sim2real."""
# High-entropy domain randomization
physics_randomizers = [
StaticFrictionRandomizer(scale=0.5), # Wide friction range
MassMultiplicationRandomizer( # Body mass variation
body_name="torso",
scale=0.2
),
JointDampingRandomizer(scale=0.3), # Damping variation
# ... more randomizers for max entropy
]
# Max-entropy RL config
entropy_coef = 0.02 # High entropy bonus
target_entropy = -4.0 # Automatic temperature tuning
def compute_entropy_metrics(self, trajectory):
"""Track entropy throughout training."""
policy_entropy = self.policy.entropy(trajectory.obs)
obs_entropy = self.estimate_obs_entropy(trajectory.obs)
return {
"policy_entropy": policy_entropy,
"observation_entropy": obs_entropy,
"entropy_ratio": policy_entropy / obs_entropy,
}
def adapt_randomization(self, real_data):
"""
Adapt domain randomization to match real robot entropy.
This is the key insight: we don't try to match exact
parameters, we match the *entropy profile*.
"""
sim_obs = self.collect_sim_observations()
real_obs = real_data.observations
# Compute entropy gap
H_sim = self.estimate_entropy(sim_obs)
H_real = self.estimate_entropy(real_obs)
if H_sim > H_real * 1.5:
# Sim too noisy, reduce randomization
self.reduce_randomization_scale(0.9)
elif H_sim < H_real * 0.7:
# Sim too deterministic, increase randomization
self.increase_randomization_scale(1.1)
# Match distribution via Wasserstein
W = self.wasserstein_distance(sim_obs, real_obs)
self.log("wasserstein_distance", W)
Why Entropy Works for Sim2Real
1. Coverage Guarantee
If policy π is optimal for ALL sims in support of p(sim),
and real world ∈ support of p(sim),
then π works in real world.
Key: Entropy maximization → widest possible support
2. Robustness via Exploration
High H(π|s) → policy doesn't overfit to single solution
→ maintains multiple viable strategies
→ can adapt when reality differs
3. Information Bottleneck
Sim and real share mutual information I(sim; real)
Maximize I → sim captures what matters about real
Ignore I → overfit to sim-specific artifacts
GF(3) Triads
entropy-sim2real (-1) ⊗ kos-firmware (+1) ⊗ mujoco-scenes (0) = 0 ✓
entropy-sim2real (-1) ⊗ jaxlife-open-ended (+1) ⊗ wobble-dynamics (0) = 0 ✓
ksim-rl (-1) ⊗ kos-firmware (+1) ⊗ entropy-sim2real (-1) = needs +1
Related Skills
ksim-rl(-1): Base RL trainingkos-firmware(+1): Deployment targetergodicity(0): Ergodic theory foundationsbirkhoff-average(-1): Time averagesfokker-planck-analyzer(-1): Distribution dynamics
References
@article{haarnoja2018sac,
title={Soft Actor-Critic: Off-Policy Maximum Entropy Deep RL},
author={Haarnoja, Tuomas and others},
journal={ICML},
year={2018}
}
@article{tobin2017domain,
title={Domain Randomization for Transferring Deep Neural Networks},
author={Tobin, Josh and others},
journal={IROS},
year={2017}
}
@article{zhao2020sim,
title={Sim-to-Real Transfer in Deep Reinforcement Learning},
author={Zhao, Wenshuai and others},
journal={IEEE TNNLS},
year={2020}
}