GRiD: GPU-Accelerated Rigid Body                  Place Zoom 
     Dynamics with Analytical Gradients              Headshot Here
                    GRiD                          Outputs
            Inputs              RBDReference    Validated Outputs
            User’s  URDFParser
            URDF                                Optimized CUDA 
                                GRiDCodeGenerator C++ Code
                           GRiDBenchmarks
                                                 Performance 
                                                 Benchmarks
                   1              1           1
         Brian Plancher , Sabrina M. Neuman , Radhika Ghosal ,
                          1,2            1
              Scott Kuindersma , Vijay Janapa Reddi
     1: Harvard University John A. Paulson School of Engineering and Applied Sciences, 2: Boston Dynamics
     GRiD: GPU-Accelerated Rigid Body               Place Zoom 
     Dynamics with Analytical Gradients           Headshot Here
        GRiD makes it easy to use the GPU
        with robotics algorithms that use rigid
        body dynamics and provides up to a
        7.2x speedup and maintains a 2.5x
        speedupwithI/O.
     GRiD: GPU-Accelerated Rigid Body               Place Zoom 
     Dynamics with Analytical Gradients           Headshot Here
     1. Why GPU Rigid Body Dynamics?
     2. GRiD’s Modular Design
     3. GRiD’s Optimizations
     4. Results
            Rigid Body Dynamics Gradients are a bottleneck                                                                               Place Zoom 
            for planning and control (e.g., nonlinear MPC)                                                                           Headshot Here
                                                                                                     Dynamics Gradient as a 
                                                                                                     Percent of Computation
                                                                                       100%
                                                                                         80%
                                                                                         60%                                                              30-90%
                                                                                         40%
                                                                                         20%
                                                                                           0%
                                                                                                      [1]        [2]       [3C]       [3G]
    [1] J. Carpentier and N. Mansrud,         [2] M. Neunert, et al., “Fast nonlinear Model     [3] Best end-to-end [C]PU and [G]PU option from B. Plancher and S. 
    “Analytical Derivatives of Rigid Body     Predictive Control for unified trajectory         Kuindersma, “A Performance Analysis of Parallel Differential Dynamic 
    Dynamics Algorithms,” RSS 2018            optimization and tracking,” ICRA 2016             Programming,” WAFR 2018