Sensor Placement Agile Robotics Program Review August 8
Sensor Placement Agile Robotics Program Review August 8, 2008 Matthew Walter, Michael Boulet, Luke Fletcher, Matthew Antone, Nick Roy, Seth Teller
Outline • • • Perception requirements Sensor types Candidate sensor placements Perception simulations Data collection Summary
Perception Tasks • Several roles expected of forklift sensors – Situational Awareness: supervisor “bot’s-eye” view – Navigation: path planning, terrain mapping, obstacle avoidance – Object Detection: finding and recognizing trucks, pallets, slots; estimating pose – Object Manipulation: pallet geometry estimation, tine insertion, load balancing – Safety: Shouted command recognition, seeing beyond load, detecting proximity of humans • Requires array of sensor types and placements
Perception Constraints • System complexity – Simplify hardware and perception algorithms • Physical constraints – Maximize visibility, protect sensors • Practical considerations – Subject to acceptable limits on size, weight, power, cost
Sensor Types • Exteroceptive (perceiving the world) Array Microphone FOV: 180 deg Color Digital Camera FOV: 90 deg Res: 752 x 480 pix Rate: 60 Hz SICK Laser Range Scanner Range: up to 80 m FOV: 180 deg Res: 1 deg Rate: 75 Hz Hokuyo Laser Range Scanner Range: up to 4 m FOV: 240 deg Res: 0. 36 deg Rate: 10 Hz • Proprioceptive (self movement) – GPS/IMU – Odometry – Mast and tine states – Wheel encoders • Interoceptive (internal state) – Motor, brake status – Strain gauges
Camera Placement Cameras provide 360 o situational awareness Right Front Rear Left
Supervisor Camera Views Left Front Right
Skirt Laser Placement Long-range skirt lasers assist in navigation and obstacle avoidance Scan in plane approximately parallel to the ground
Pushbroom Laser Placement Long-range pushbroom lasers assist in navigation and terrain mapping Front Top: 10 -20 m out Rear: 10 -20 m out Front Bottom: Close range, beneath load
Temporal Scan Persistence • Single scan sees only narrow “slice” of world • Each scan placed in local 3 D coordinate frame – Requires knowledge of forklift, mast, and tine pose – Allows aggregation of multiple scans over time Drive forward Raise mast
Lasers for Navigation
Pallet Sensor Layout Short-range lasers move with mast and tines to perceive pallets and slots Pushbroom (2 -5 m out) Skirt lasers (“virtual tines”)
Pallet Sensing Simulations • Goal: study effects of sensor placement on pallet and slot perception • Use system infrastructure for simulation – Virtual 3 D environment containing objects of interest – Simulated returns from lasers in various configurations – Data acquisition with different pallet types, pallet poses, ranges, approach maneuvers
Pallet Sensing Simulations
Simulated Data: Ground Pallet No mast or tine motion Approaching Pallet on ground Raise/lower mast Tilt mast
Simulated Data: Truck Pallet Approaching Pallet on truck Mid-field, ~4 m Far-field, ~18 m Near-field, ~2 m
Real Data Collection • Goal: acquire representative sensor data • Pallet approach and slot detection • Mast and tine movement to ‘scan’ objects • Realistic non-level, non-smooth surfaces
Data Collection Sensor Layout Camera Pushbroom Vertical Skirt
Short Range Laser Scans
Longer Range Laser Scans
Summary • Sensor types determined to meet operational needs – Support situational awareness, navigation, object detection, object manipulation, safety tasks • Studies performed using simulated and real data sets – 3 D perception simulator with different configurations – Data collected from real forklift on realistic terrain – Objects ‘scanned’ using mast motion • Additional studies to be performed in coming weeks
Timeline (Year 1) start of year 1 apr 08 end of year 1 today sep 08 nov 08 perception simulator preliminary placement studies navigation sensors on forklift experimental configurations all sensors on forklift final sensor configuration precise body/mast calibration experimental validation additional simulation investigation of year 2 sensors jan 09 mar 09
- Slides: 22