+
+
+
+
+
+
+
+
+
+
+ Abstract
+
+ Contact estimation is a key ability for limbed robots, where making and breaking contacts has a direct impact on state estimation and balance control. Existing approaches typically rely on gate-cycle priors or designated contact sensors. We design a contact estimator that is suitable for the emerging wheeled-biped robot types that do not have these features. To this end, we propose a Bayes filter in which update steps are learned from real-robot torque measurements while prediction steps rely on inertial measurements. We evaluate this approach in extensive real-robot and simulation experiments. Our method achieves better performance while being considerably more sample efficient than a comparable deep-learning baseline.
+
+
+
+
+
+
+ Results
+
+
+
+
+
+
+
+ 
+
+
+
+ We implement a Bayesian filter to estimate the contact probability at each timestep by fusing IMU readings and joint torque measurements. We infer transition probabilities through spectral analysis, and use the torque estimates to infer measurement probabilities.
+
+ +
+
+
+
+
+ 
+
+
+
+
+ We implement and test the performance of the estimation filter to detect contacts on the wheeled-biped Upkie. Our study focuses on drops from unforeseen steps, where the robot alternates between contact and free-fall phases. We set up a test course where the robot travels over a horizontal floor, climbs a meter-long ramp with an 8° incline, drops 3 cm from the ramp onto a platform, rolls forward and + drops again from a 12-cm step, which we repeated for a total number of 10 drops or 5 laps. We see that our method achieves a 100% success rate in detecting the contact events, with better precision than the two baselines. +
+ + We implement and test the performance of the estimation filter to detect contacts on the wheeled-biped Upkie. Our study focuses on drops from unforeseen steps, where the robot alternates between contact and free-fall phases. We set up a test course where the robot travels over a horizontal floor, climbs a meter-long ramp with an 8° incline, drops 3 cm from the ramp onto a platform, rolls forward and + drops again from a 12-cm step, which we repeated for a total number of 10 drops or 5 laps. We see that our method achieves a 100% success rate in detecting the contact events, with better precision than the two baselines. +
+ 
+
+
+
+
+ We see that our method achieves a 100% success rate in detecting the contact events, with better precision than the two baselines, while maintaining competitive latencies. +
+ + We see that our method achieves a 100% success rate in detecting the contact events, with better precision than the two baselines, while maintaining competitive latencies. +
+ 
+
+
+