Drifting MPC: Trajectory Optimization Without a Simulator Using Offline Dataset Learning

A new framework called Drifting MPC solves a fundamental problem in robotics and control: how to plan optimal trajectories when you don't have a simulator and can't model the system dynamics. The answer: learn from existing trajectory data alone.

The Challenge

Traditional model predictive control (MPC) requires:

• A dynamics model — Mathematical description of how the system evolves
• A simulator — Ability to simulate trajectories before executing them
• Online computation — Real-time optimization at each planning step

But what if:

• You don't know the dynamics?
• You can't simulate trajectories?
• You only have a dataset of past trajectories?

Drifting MPC Approach

The framework combines:

1. Drifting generative models — Learn a conditional distribution over trajectories from offline data
2. Receding-horizon planning — Plan ahead while adapting as new information arrives
3. Optimality-data tradeoff — Balance between generating optimal plans and staying close to demonstrated behavior

Key Properties

• Unique solution — The resulting distribution is the unique solution of an objective trading off optimality with closeness to offline prior
• One-step inference — Near-optimal trajectories with single-step generation efficiency
• No simulation required — Works directly from offline trajectory data

Applications

• Robotics — Plan robot motions without precise physics models
• Autonomous vehicles — Learn driving behavior from demonstration data
• Manufacturing — Optimize industrial processes from historical trajectory data
• Any domain — Where you have trajectory data but no simulator