Localization for the Naos Brian Coltin, Joydeep Biswas, Sooho
SPL Soccer – Nao Robots • Sensors – Camera – Referee client • Computation – x86 AMD GEODE 500 MHz • Playing “reasonable soccer” requires: – Speed
SPL Game Field
Tradeo fg s • Speed of Computation: – Simplified models – Fewer particles ⇒ Inaccuracy • Speed of Convergence ⇒ Loss of diversity • Accuracy – Accurate models – More hypotheses ⇒ Slow computation
Current Method • Very large number of particles (worst case) • Sensor Resetting • Adaptive Importance Sampling
Proposed Approach • Can we do better if the sensor model includes a “correctional gradient”? • Questions: – Will this work? – Will this prematurely get rid of diversity? – Will it provide a noticeable improvement? • Increased accuracy • Faster convergence – Will it be robust to false positives?
Related Work • Hybrid Monte Carlo [Duane, AD et al., Physics letters B, 1987] • Rao Blackwellisation (?)
Recommend
More recommend
