INSAIT Introduces DiffSim Trinity: A New Paradigm for Autonomous Driving Based on Differentiable Simulation

INSAIT, part of Sofia University “St. Kliment Ohridski”, introduces DiffSim Trinity (diffsimtrinity.insait.ai), a new paradigm for autonomous driving built on differentiable simulation. The work builds on Waymax, the differentiable simulator developed by Waymo, a global leader in autonomous driving that originated from Google. DiffSim Trinity unifies the full sensor → action → outcome pipeline, enabling more precise, robust, and reliable autonomous driving systems.

Modern end-to-end autonomous driving models are trained to map sensor inputs directly to driving actions such as steering and acceleration. However, the physical dynamics of the world – how those actions affect vehicle motion and the surrounding environment—are typically treated as a black box. Differentiable simulation addresses this limitation by allowing learning signals to pass through physical world dynamics. This enables models to learn not only which actions to take, but how actions lead to outcomes, supporting data-efficient training, direct optimization for safety and comfort, and reasoning about possible future events.

DiffSim Trinity translates this idea into practice through three research works accepted at leading machine learning and robotics venues, each highlighting a key capability enabled by differentiable simulation.

For control, the project presents the first fully end-to-end driving policy trained with differentiable simulation on the Waymo Open Motion Dataset, demonstrating how physical dynamics can be directly incorporated into policy learning.

For planning, DiffSim Trinity introduces the first world-modeling algorithms for autonomous driving based on differentiable simulation. These methods enable counterfactual “what-if” reasoning, allowing a vehicle to predict how different actions could change future traffic scenarios – an essential requirement for safer autonomous driving.

For search, the work showcases how differentiable simulation can be used not only during training, but also for online reasoning at deployment time. In this setting, the vehicle uses imagined future outcomes to efficiently search for and optimize the best possible trajectory given expected changes in its surroundings.

Altogether, DiffSim Trinity demonstrates the value of a unified, end-to-end pipeline from sensors to outcomes, enabling autonomous driving systems that are safer, more comfortable, and more humanlike in complex and dynamic environments.

The DiffSim Trinity was developed in collaboration with the University of Zurich and ETH Zürich, and the three works were published at top international venues in robotics and artificial intelligence, including IROS 2025 and AAAI 2026.

INSAIT congratulates all contributors: Asen Nachkov, Jan-Nico Zaech, Danda Pani Paudel, Xi Wang, Davide Scaramuzza, and Luc Van Gool.