Fault detection and mitigation from uninterpreted data of robotic - PowerPoint PPT Presentation

1 Fault detection and mitigation from uninterpreted data of robotic sensorimotor cascades Andrea Censi * Magnus Håkansson # Richard M. Murray Caltech LTH Caltech * 5th year graduate student, # spent a California summer defending in a few weeks, down in the basement currently looking for a job collecting data Outline 1. Robust agents need to learn/verify the models that they use. 2. Bootstrapping: learning low-level models for robotic sensorimotor cascades 3. “Sensorimotor faults” can be defined independently of a nominal model. 4. “Orthogonal” integration within a traditional robot architecture

2 • Models used by robotic agents are often implicit and never verified. Sutton’s “Verification principle”: An intelligent system can create and maintain knowledge only to the extent that it can verify that knowledge itself. • How to predict the unpredictable? miscalibration interference sabotage hardware failures

3 • Fault detection is an estimation problem, but what should be the design goal? accuracy * (robustness to assumptions) accuracy max max (design effort) * computation a simpler approach Bayesian filters based on low-level ✓ the optimal solution sensorimotor models × not optimal × sensitive to prior assumptions ✓ fewer assumptions × needs lots of design effort that could be violated ✓ “surprising” effects p (bird) with reduced design effort p (image|bird droppings) p (image|¬ bird droppings)

4 Outline 1. Robust agents need to learn/verify • Bootstrapping = starting the models that they use. from no prior information 2. Bootstrapping: learning low-level • All robots << all systems models for robotic sensorimotor cascades • Sensors are more similar 3. “Sensorimotor faults” can be defined than what you would think. independently of a nominal model. • Models for low-level 4. “Orthogonal” integration within sensorimotor learning a traditional robot architecture

5 • Usually agents start with a model (implicit or explicit) of the robot and the physical environment. model physical environment agent observations commands “world” or “sensorimotor cascade” physical environment robot

6 • In the bootstrapping scenario, ✓ zero assumptions the agent has no prior knowledge that can be violated about its sensors, its actuators, and the external environment. ? bootstrapping agent uninterpreted uninterpreted observations commands “sensels”: “world” or “sensorimotor cascade” pixels, range readings, unknown external unknown … actuator(s) sensor(s) environment

7 • The “set of all robots” is much smaller than the set of all dynamical systems. all dynamical systems all robots

8 • “Canonical robotic sensors” [ICRA’11] have similar dynamics at the sensel level. y ( s ) = intensity at point s field-sampler exactly bilinear range-finder y ( s ) = distance in direction s bilinear, up to a nonlinearity camera y ( s ) = luminance in direction s bilinear, up to a hidden state v ( t ) , ω ( t ) : kinematic velocities

9 all dynamical systems • What are the simplest models that we can use to detect faults? all robots more keeps priors accurate on bird migration patterns instantaneous models today of how the commands u ( t ) determine y ( t+ d t ) too simple for discriminating faults simpler

10 all dynamical systems • Several classes of models for low-level sensorimotor data all robots have been studied. simpler BDS Models a bilinear relation between ẏ , y , and u : less assumptions more data required [ICRA’11] BGDS Models bilinear flows of the observations space: [IROS’11] DDS Models diffeomorphisms of the observations space. more complex Paper ThA01.2, Thursday more assumptions 08:45 − 09:00, Meeting Room 1 (Mini-sota) less data required

11 Outline 1. Robust agents need to learn/verify the models that they use. • Traditional fault detection requires a nominal model 2. Bootstrapping: learning low-level models for robotic sensorimotor cascades • “Sensorimotor faults”: “faulty”= uninformative 3. “Sensorimotor faults” can be defined = unpredictable independently of a nominal model. • Application to camera data 4. “Orthogonal” integration within a traditional robot architecture

12 • Traditional fault detection requires a nominal model. Traditional fault detection given 1. Acquire a nominal model identified / learned for the healthy system. residuals 2. Check the observed data likelihood ratios against the model. shut off the reactor 3. Do something if the data does not support the model. ...

13 • “Sensorimotor faults” can be defined without reference to a nominal model. a “faulty” sensel y i does not give information about u . DEF: a “faulty” actuator u j does not give information about y . ... y a u a u y commands observations ... ... y b u b “faulty” sensels “faulty” actuators examples: a dead pixel a random pixel an occluded pixel

14 • “Sensorimotor faults” can be defined without reference to a nominal model. a “faulty” sensel y i does not give information about u . DEF: ... y a u a u y commands observations ... ... y b u b “faulty” sensels distance between p ( u t | y :t ) “ usefulness” of the i- th sensel and p ( u t | y :t - { y i }) = correlation between the observed ẏ assuming additive noise, for a limited class of models, and the prediction of the learned model and a particular distance (“ predictability ”)

15 “Sensorimotor faults” detection Traditional fault detection 1. Learn from the running system 1. Acquire a nominal model from sensorimotor data ( u ( t ), y ( t )) for the healthy system. (for some model class C). 2. Check the observed data 2. Compute the sensel “usefulness”: against the model. distance between p ( u t | y :t , C ) 3. ... and p ( u t | y :t - { y i }, C ). 3. “Useless” sensels are marked as “faulty”. 4. ...

16 • Differential-drive robot equipped with two cameras Agent fits the model: [IROS’11] “Sensorimotor faults” detection 1. Learn from the running system from sensorimotor data ( u ( t ), y ( t )) (for some model class C). y ( t ) 2. Compute the sensel “usefulness”: ER1 distance between p ( u t | y :t , C ) and p ( u t | y :t - { y i }, C ). 3. “Useless” sensels are marked as “faulty”. 4. ... a second camera u ( t ) : linear/angular looks through a mirror velocities

17 ✓ with no prior knowledge • “Faults” detected: about the robot • self-occlusions ✓ with minimal computation • border between images (learning/inference) • sampling limitations (points at infinity) Usefulness / predictability useless / useful / unpredictable predictable y ( t ) (“faulty”) ER1 robot frame border and between wheels images a second camera u ( t ) : linear/angular looks through a mirror velocities

18 ✓ with no prior knowledge • No “faulty” actuators in this example, about the robot but the instantaneous commands ✓ with minimal computation anomaly signal can detect: (learning/inference) • Violations of planarity assumption • Delays in commands execution • Synchronization issues Instantaneous anomaly signal for u ( t ) robot drives delays over a bump delays … and off the bump synchronization issues time (s)

19 Outline 1. Robust agents need to learn/verify the models that they use. • How to integrate 2. Bootstrapping: learning low-level these techniques models for robotic sensorimotor cascades into an existing architecture, 3. “Sensorimotor faults” can be defined with minimal changes? independently of a nominal model. • Application 4. “Orthogonal” integration within to range-finder data a traditional robot architecture

20 • How to integrate these techniques in a traditional architecture with minimal changes ? traditional traditional agent agent panic sanitized y signal bootstrapping agent y u u glue sensel usefulness u y y robot robot • A bootstrapping agent computes the sensel usefulness. • A glue component marks useless sensels as invalid. • ...with an adaptive threshold on the basis of a “panic” signal sent by the traditional agent (no tuning needed)

21 • Landroid robot equipped with a Sensel usefulness / predictability range finger, partially occluded by (vibrating) WiFi antennas antenna antenna antennas range finder threshold α sensel # explorer u ( t ) panic y’ signal panic mode normal operation sensel usefulness bootstrapping α agent u glue sensels u y y marked sensel invalid s Landroid y (s, t )

21 • Landroid robot equipped with a Sensel usefulness / predictability range finger, partially occluded by (vibrating) WiFi antennas antenna antenna antennas threshold α range finder sensel # explorer u ( t ) panic y’ signal panic mode normal operation sensel usefulness bootstrapping α agent u glue sensels u y y marked sensel invalid s Landroid y (s, t )