Główne pojęcia
Learning diffusion models as priors accelerates robot motion planning optimization by sampling from posterior trajectory distributions.
Statystyki
"Our main contributions are: (1) we learn a trajectory generative model with a diffusion model using expert trajectories generated with an optimal motion planning algorithm; (2) we formulate the motion planning problem as planning-as-inference by sampling from a posterior distribution leveraging guidance in diffusion models; (3) to validate our approach we present results in several environments with increasing difficulty; (4) we empirically demonstrate that learning and sampling from the diffusion model speeds up motion planning without an informed prior and is better than a commonly used generative model."
"The trajectory cost is cEE(τ) = PH−1 k=0 dSE(3) ((T (q[k]), Tg[k])), where T (q[k]) is the end-effector pose at waypoint k and Tg[k] the desired one."
"The diffusion models are trained for 25 diffusion steps with exponential scheduling (we found it to work better than linear scheduling [24])."
Cytaty
"Our main contributions are: (1) we learn a trajectory generative model with a diffusion model using expert trajectories generated with an optimal motion planning algorithm; (2) we formulate the motion planning problem as planning-as-inference by sampling from a posterior distribution leveraging guidance in diffusion models; (3) to validate our approach we present results in several environments with increasing difficulty; (4) we empirically demonstrate that learning and sampling from the diffusion model speeds up motion planning without an informed prior and is better than a commonly used generative model."