Dm Control Prediction Eval

dm-control-prediction-eval

Neural Motion Simulator: Pushing the Limit of World Models in Reinforcement Learning — Hao et al. (2025) (arXiv:2504.07095, 2025)

What this evaluates

Evaluates long-horizon physical state prediction accuracy of a neural motion simulator in continuous control environments, and measures its effectiveness for zero-shot reinforcement learning by comparing prediction horizons and minimal training step requirements.

Datasets

• DM Control — total ?; splits: test (-1)

Metrics

• MSE loss (primary) — range: other
  • Mean Squared Error between predicted and ground-truth future states over a specified prediction horizon.
• prediction horizon — range: other
  • The number of prediction steps at which MoSim achieves the same MSE loss as DreamerV3 at a fixed 16-step horizon.

Input / output format

Input: Current environment state and action sequence in DM Control physics simulation.

Output: Predicted future state vectors over a multi-step horizon.

Scoring recipe

def compute_mse(pred, gold):
    return np.mean((pred - gold) ** 2)

def evaluate_horizon(model, env, baseline_mse):
    for t in range(1, 100):
        pred = model.predict(env.state, env.actions, steps=t)
        gold = env.get_ground_truth(t)
        if compute_mse(pred, gold) <= baseline_mse:
            return t
    return -1

Common pitfalls

• Confusing control steps (physics timesteps) with action repeats; evaluation must account for environment-specific control step multipliers.
• Training TD-MPC2 on random data produces a meaningless latent space; only provided checkpoints should be used for fair comparison.
• Prediction horizon is relative to a fixed 16-step DreamerV3 baseline, not an absolute error threshold.

Evidence (verbatim from paper)

Table IV uses the MSE loss of DreamerV3 at a 16-step prediction horizon as a reference. It presents the prediction horizon at which MoSim achieves the same loss, providing a more intuitive measure of MoSim's predictive performance.

Citation

@misc{hao2025neuralmotionsimulator,
  title={Neural Motion Simulator: Pushing the Limit of World Models in Reinforcement Learning},
  author={Hao et al. (2025)},
  year={2025},
  note={arXiv:2504.07095}
}

• arXiv: 2504.07095