SIMULATION DISTILLATION

Pretraining World Models in Simulation for Rapid Real-World Adaptation

Jacob Levy*1, Tyler Westenbroek*2, Kevin Huang2, Fernando Palafox1, Patrick Yin2, Shayegan Omidshafiei3, Dong-Ki Kim3, Abhishek Gupta2, David Fridovich-Keil1

1 UT Austin 2 UW 3 FieldAI * Equal Contribution

Rapid Real-World Adaptation

Simulation Distillation (SimDist) rapidly overcomes the sim-to-real dynamics gap through adaptation in the real world, resulting in substantial gains in task execution on both precise manipulation and quadrupedal locomotion tasks.

Peg Insertion

Zero-shot: 25% success
After 60 trials: 85% success

Table Leg

Zero-shot: 10% success
After 80 trials: 85% success

Slippery Slope

Zero-shot: 0% success
After 36 minutes: 93% success

Foam

Zero-shot: 40% success
After 32 minutes: 100% success

Why World Models?

End-to-end RL finetuning in the real world often forgets useful priors from simulation. SimDist keeps global task structure fixed and updates only what changes most across domains: dynamics.

End-to-end RL Finetuning is Hard

Sim-to-real policies fail under dynamics mismatch. Standard RL finetuning entangles representation, dynamics, and returns, forcing relearning of the entire task structure in the new domain.

World Models Factorize Problem Structure

World model architecture showing encoder, latent dynamics, and value components

World models decompose task structure in a modular format we can exploit for efficient real-world adaptation.

Leveraging Simulation for Transferable Priors

Domain randomization and privileged state supervision yield robust representation and reward/value models in simulation, allowing real-world finetuning to focus solely on dynamics adaptation.

Introducing Simulation Distillation

Simulation Distillation (SimDist) is a scalable framework that distills structural priors from a simulator into a latent world model and enables rapid real-world adaptation via online planning and supervised dynamics finetuning.

Hover over a box to view its detail below.

Real-World Results

Success rate for two manipulation tasks, computed over 20 trials, and average forward progress for two quadruped locomotion tasks, averaged across all 15 trials (3 speeds, 5 trials each), as a function of real-world finetuning data. For manipulation, we consider two difficulties: initial conditions drawn from a Narrow or Wide grid.

SimDist exhibits rapid and consistent improvement with limited data by finetuning only the latent dynamics model while planning with frozen reward and value models. In contrast, direct policy finetuning with the baselines shows limited or no improvement under the same data budgets.

Hover a legend item to highlight a curve.
Click to toggle it on/off.

Hover a plot to preview its video. Click to expand.

Static fallback plot for real-world results

Interactive charts load here from assets/data/results.json.

Successful Return Model Transfer

The planner used by SimDist can only improve behavior if it can reliably distinguish trajectories with high and low returns. This requires both accurate dynamics prediction and successful transfer of reward and value models. We examine value transfer below, which plots predicted values over time for successful and failed trajectories.

For the successful rollout, predicted value increases consistently over time, while remaining lower for the failed trajectory. Thus the transfered encoder and value function can reliably discriminate between successful and failed trajectories.

Hover over the plot to view the associated video frame.

Success
Failure

Improving Dynamics Prediction with Finetuning

Next, we examine the effect finetuning has on dynamics prediction accuracy. Adapting the dynamics model is essential for effective planning, as both the reward and value estimates are computed over predicted trajectories.

Finetuning drastically lowers dynamics prediction loss for a held out quadruped slippery slope trajectory.

During this trajectory, the front-left foot slips.

Foot prediction comparison visualization during slip event

At this same instant, the finetuned model correctly anticipates the future slippage, while the pretrained model fails to do so.

Planning Under Adapted Dynamics

Because we finetune only the dynamics, adaptation is fast: improved predictions immediately reshape planning behavior, driving the performance gains seen in our results.

At each timestep, the planner queries the adapted world model iteratively to close in on the optimal trajectory.
Slippery Slope: trajectory samples generated with the finetuned model reflect the real-world contact dynamics and lead the planner to select plans that account for foot slip.
Foam: the finetuned model produces plans that account for real world dynamics mismatch, while the pretrained model generates qualitatively different trajectories