Event

Shanghai Jiao Tong University team proposes DiffDreamer quadrotor control method

1 reports · 1 sources

The Zou Danping team at Shanghai Jiao Tong University proposed DiffDreamer, which combines a learned world model with differentiable simulation for agile quadrotor control; the paper has been accepted by RAL 2026. The method has the world model predict external dynamics and the differentiable simulation propagate the robot's physical state; in the aggressive tracking task, the distance error is 0.381 m, versus 0.793 m for the strongest baseline; in dynamic gate crossing, at the three gate speeds of 1.0/1.5/2.0 m/s, the number of gates passed is 16/12/8, versus 12/4/1 for MPC. These results are based on external state feedback provided by motion capture; the tracking error is "mean ± standard deviation" and relies on a calibrated differentiable simulation and external ground truth, and cannot be equated with open-scene performance under noisy perception. The experiments were completed on a self-developed quadrotor with 3.5-inch propellers, an onboard Jetson Orin NX, and a Betaflight flight controller, with the policy running onboard at over 100 Hz; the paper explicitly defines it as a proof of concept.