Zürich Autonomous Systems Lab Cedric Pradalier [email protected] ICRA Workshop on Planetary Rovers, May 2010
Mar 27, 2015
Zürich
Auto
nom
ous
Syst
em
s La
b2
Welcome to Anchorage
Zürich
Auto
nom
ous
Syst
em
s La
b3
Outline Autonomous Systems Lab
Brief summary of the space-related activities
Hardware platforms◦ Eurobot EGP Prototype◦ ExoMars breadboard
Embedded Software◦ Lowering friction requirements using optimised
torque distribution◦ Learning what’s come ahead
Zürich
Auto
nom
ous
Syst
em
s La
b4
Lab of Pr. Siegwart◦ www.asl.ethz.ch◦ ETH Zürich – Switzerland◦ 20 PhD / 40 Total
Education◦ Lectures: Bachelor / Master◦ Project supervision
Research◦ Vision: Create machines that know what they do◦ Three research line:
The design of robotic and mechatronic systems Navigation and mapping Product design methodologies and innovation
Autonomous Systems Lab
ZürichAutonomous Systems Lab
Overview, Crab, Eurobot EGP Prototype
Exomars Breadboard
Zürich
Auto
nom
ous
Syst
em
s La
b6
◦ Micro Air Vehicles◦ Walking and Running
Quadruped Robots◦ Service Robots◦ Autonomous Robots/Cars
for Inner City Environments◦ Inspection Robots◦ Space Robots for Planetary
Exploration◦ Autonomous sailing/electric
boats
ASL – ETH Zurich
Zürich
Auto
nom
ous
Syst
em
s La
b7
Nanokhod
Shrimp & Solero◦ Passive suspension
systems◦ 6 motorized wheels◦ 2 steering
◦ Very good terrainability!
ASL rovers background
Zürich
Auto
nom
ous
Syst
em
s La
b8
RCL-E
RCL-C
CRAB
Exomars: Pre-study phase A
Zürich
Auto
nom
ous
Syst
em
s La
b9
Platform◦ Passive suspension◦ 6 Motorized wheels◦ 4 Steering
Mobile robots◦ Confronted to environments
which are unknown◦ Difficulty to: Model before-hand the
environment of the rover. Predict its terrain interaction
characteristics.
CRAB rover
Zürich
Auto
nom
ous
Syst
em
s La
b10
ExoMars Breadboard
Zürich
Auto
nom
ous
Syst
em
s La
b11
ExoMars Breadboard
Zürich
Auto
nom
ous
Syst
em
s La
b12
Authorization denied…
Test plan and results
Zürich
Auto
nom
ous
Syst
em
s La
b13
Eurobot: ◦ Multi-arm astronaut assistant◦ Developed by Thales (and others?) for ESA
EGP = Eurobot Ground Prototype◦ Put some wheels and perception under the Eurobot◦ Experiment on the concept of an astronaut assistant
EGP Rover Prototype
Picture from Didot et al. IROS’07
Zürich
Auto
nom
ous
Syst
em
s La
b14
Ability to carry and power Eurobot (150Kg)
Ability to transport an astronaut in full EVA (100Kg)
Power autonomy for multiple hours, fast recharge
◦ 150kg of lead-acid batteries
Ability to perceive its surrounding, plan path, follow
an astronaut, using a stereo-pair
Rough terrain capabilities (15 deg slopes, 15cm
steps)
Cheap !!!
EGP Rover – Requirements
Zürich
Auto
nom
ous
Syst
em
s La
b15
Mechanical design
Zürich
Auto
nom
ous
Syst
em
s La
b16
Mechanical design
Zürich
Auto
nom
ous
Syst
em
s La
b17
Implementation
Zürich
Auto
nom
ous
Syst
em
s La
b18
Suspension
Zürich
Auto
nom
ous
Syst
em
s La
b19
Integration
880kg, without astronaut…
Zürich
Auto
nom
ous
Syst
em
s La
b20
Integration
ZürichAutonomous Systems Lab
Optimised Torque ControlLearning what comes ahead
Zürich
Auto
nom
ous
Syst
em
s La
b22
Optimised torque control Principle
◦ It is possible to put more torque on wheel with more load
Requirements◦ Measurement of contact
point on each wheel
◦ Static model to deduce the wheel load from the contact points and the rover state
Results submitted to IROS’10
Zürich
Auto
nom
ous
Syst
em
s La
b23
Control loop
Zürich
Auto
nom
ous
Syst
em
s La
b24
Test setup and hardware
Zürich
Auto
nom
ous
Syst
em
s La
b25
Results
Zürich
Auto
nom
ous
Syst
em
s La
b26
Results
ZürichAutonomous Systems Lab
Ambroise [email protected]
Zürich
Auto
nom
ous
Syst
em
s La
b29
Two types of sensors needed◦ Remote sensors → Remote Terrain Perception data◦ Local sensors → Rover-Terrain Interaction
data Data association Prediction
◦ What are the Rover-Terrain Interaction characteristics?
Approach: Basic concept
?
Zürich
Auto
nom
ous
Syst
em
s La
b30
Delay
Approach: Architecture overview RTILE Rover-Terrain Interactions Learned from
Experiments
SOFTWARE
HARDWARE Actuators
Controller
Path Planning Prediction
Learning
Database
ProBT
Near to far
Local Sensors Remote Sensors
Obst. Det.
Trafficability & Terrainability
Traversability
Zürich
Auto
nom
ous
Syst
em
s La
b31
Outline
SOFTWARE
HARDWARE Actuators
Controller
Path Planning Prediction
Learning
Database
ProBT
Near to farDelay
Local Sensors Remote Sensors
Obst. Det.
Zürich
Auto
nom
ous
Syst
em
s La
b32
Data acquisition: 2D example
Grid based approach◦ Remote Image acquisition◦ Local Position of the wheels◦ Samples When learning occurs
Near to far
Samples can be used for the learning mechanism.
Remote
Local
Features association
Zürich
Auto
nom
ous
Syst
em
s La
b33
Bayesian model Goal
◦ Local features predicted based on remote features
Bayesian model◦ Joint distribution and decomposition◦ Introduce abstraction classes and
Question
→
Class association
Local classification
Remote classification
Zürich
Auto
nom
ous
Syst
em
s La
b37
Outline
SOFTWARE
HARDWARE Actuators
Controller
Path Planning Prediction
Learning
Database
ProBT
Near to farDelay
Local Sensors Remote Sensors
Obst. Det.
Zürich
Auto
nom
ous
Syst
em
s La
b38
Prediction Process
Remote Subspace Local Subspace
Fr = 0.5
Prediction
20% 50% 30%
Zürich
Auto
nom
ous
Syst
em
s La
b39
Path planner – E*◦ Wavefront propagation
Navigation function◦ Gradient descent◦ Propagation cost
Process
Adaptive navigation
assumption T = 1
Image acquisition
Fl prediction Propagation costs
Zürich
Auto
nom
ous
Syst
em
s La
b40
Outline
Zürich
Auto
nom
ous
Syst
em
s La
b41
Rover-Terrain Interaction metric◦ ◦ The smaller, the better
Remote feature space◦ Camera◦ Color description
Trajectory adaptation
Absolute cost method◦ Idea of tradeoff between
What can be gained in terms of , meaning The deviation it imposes from the default trajectory
◦ Dynamically adapts to the terrain representation
Propagation costs function
Very bad
Very good
Good
Start Goal
?
Zürich
Auto
nom
ous
Syst
em
s La
b42
RTILE: Results Adaptive navigation Test environment in Fluntern
◦ 3 terrains Grass softest (best) Tartan Asphalt hardest (worst)
•Automatically driven
•6 cm/s
•No prior
•Learning every 6 m
Zürich
Auto
nom
ous
Syst
em
s La
b43
RTILE: Results “complete” Test of the complete approach
Waypoints x [m] y [m]
0 0.0 0.0
1 15.0 0.0
2 15.0 -15.0
3 0.0 -15.0
4 0.0 -2.5
5 12.5 -2.5
6 12.5 -15.0
Zürich
Auto
nom
ous
Syst
em
s La
b45
Summary RTILE: Rover-Terrain Interactions Learned from
Experiments◦ End-to-end approach
Online learning Navigation adapted accordingly Integrated within the CRAB platform
◦ Tradeoff distance vs MRTI
20% MRTI improvement 10% longer distance
◦ Terrain description Consistent interaction with E* Dynamical adaptation of the propagation costs
RTILE improves the rover behavior
Zürich
Auto
nom
ous
Syst
em
s La
b46
Future work Improvements
◦ Add feature spaces (subspaces) for a better terrain description
◦ Use additional sensors Local: Tactile wheels, Microphones, and so on… Remote: Google earth map (increase FOV), Lidar
◦ Improved features Remote: Fourier based, Co-occurrence matrix, and
so on…◦ Learning
Clustering step (GWR)
Outlook◦ Energetic description◦ Learn as well the behavior of the rover
Zürich
Auto
nom
ous
Syst
em
s La
b47
Questions?