Thank you to all speakers, participants and contributors and our sponsor Nvidia for making this workshop a great success!

All talk recordings are available on YouTube.

Differentiable optimization plays a key role in connecting machine-learning frameworks to model-based approaches. It enables the backpropagation of gradient information resulting from solving optimization problems, such as those that arise when simulating a robot interacting with its environment. 

The promise of differentiable simulation is that the development and dissemination of mature tools, similar to those developed for vision and language in deep learning (e.g., PyTorch, Jax, TensorFlow, etc.), will open the door to seamlessly complementing existing physical models with data from the real world or to using gradients from simulation or estimation to train control policies efficiently. However, computing well-behaved gradients in settings with physical interactions is challenging because of the inherent non-smoothness of contact dynamics. Similarly, non-smooth operations such as rasterization can make common sensors in robotics such as cameras non-differentiable by design. 

The workshop is framed around three main subtopics:

Differentiable optimization and simulation
Algorithms and open-source frameworks targeted to provide gradients of non-smooth problems.

Improving simulations from data
Methods and examples for improving models based on real-world measurements through backpropagation.

Applications in policy learning and control
Application of differentiable optimization to increase efficiency and performance of policy learning.

This workshop aims to bring together academic and industry researchers and practitioners to highlight the current challenges and developments of differentiable simulation and discuss the field's practical applications, future directions, and limitations. In particular, our primary target audience is researchers from both the robotics and the learning communities working on model-based and end-to-end learned methods, and anywhere in between, to enable cross-pollination between these two complementary approaches, as enabled by differentiable optimization. Similarly, we draw speakers and panelists from both fields and emphasize speakers from industry to encourage the practical relevance of discussions and foster communication between academia and industry.