[2505.03296] The Unreasonable Effectiveness of Discrete-Time Gaussian Course of Mixtures for Robotic Coverage Studying

View a PDF of the paper titled The Unreasonable Effectiveness of Discrete-Time Gaussian Course of Mixtures for Robotic Coverage Studying, by Jan Ole von Hartz and three different authors

View PDF
HTML (experimental)

Summary:We current Combination of Discrete-time Gaussian Processes (MiDiGap), a novel strategy for versatile coverage illustration and imitation studying in robotic manipulation. MiDiGap allows studying from as few as 5 demonstrations utilizing solely digicam observations and generalizes throughout a variety of difficult duties. It excels at long-horizon behaviors akin to making espresso, extremely constrained motions akin to opening doorways, dynamic actions akin to scooping with a spatula, and multimodal duties akin to hanging a mug. MiDiGap learns these duties on a CPU in lower than a minute and scales linearly to giant datasets. We additionally develop a wealthy suite of instruments for inference-time steering utilizing proof akin to collision indicators and robotic kinematic constraints. This steering allows novel generalization capabilities, together with impediment avoidance and cross-embodiment coverage switch. MiDiGap achieves state-of-the-art efficiency on numerous few-shot manipulation benchmarks. On constrained RLBench duties, it improves coverage success by 76 proportion factors and reduces trajectory value by 67%. On multimodal duties, it improves coverage success by 48 proportion factors and will increase pattern effectivity by an element of 20. In cross-embodiment switch, it greater than doubles coverage success. We make the code publicly accessible at this https URL.