is motion planning overrated? - personal robotics lab...inverse dynamics controller robot torques...
TRANSCRIPT
Is Motion Planning Overrated?Jeannette Bohg - Interactive Perception and Robot Learning Lab - Stanford
Is Motion Planning Overrated?Jeannette Bohg - Interactive Perception and Robot Learning Lab - Stanford
tl;dr; Yes
Is Motion Planning Overrated?Jeannette Bohg - Interactive Perception and Robot Learning Lab - Stanford
Is Motion Planning Overrated?Jeannette Bohg - Interactive Perception and Robot Learning Lab - Stanford
tl;dr; It depends
Robot System DesignDynamic and Uncertain Environment
Robot System DesignDynamic and Uncertain Environment
Robot System DesignDynamic and Uncertain Environment
Real-Time Perception meets Reactive Motion Generation
Inverse Dynamics Controller
RobotTorques
Joint Sensors
Robot State Acceleration Policies
Camera
World
Interaction
World Model
Object Tracker
Arm Tracker
Motion Optimizer
Kappler et al. Real-Time Perception meets Reactive Motion Generation. RA-L + ICRA’18. Finalist 2018 Amazon Best Systems Paper
One time step in the system
One time step in the system
Raw Sensory Data
30Hz-1kHz
One time step in the system
Raw Sensory Data Processed Sensory Data
30Hz-1kHz 15-30Hz
One time step in the system
Raw Sensory Data Processed Sensory Data Local and optimised policies
30Hz-1kHz 15-30Hz 30kHz 5-10Hz
One time step in the system
Raw Sensory Data Processed Sensory Data Local and optimised policies Fused Policy
30Hz-1kHz 15-30Hz 30kHz 5-10Hz 30 kHz
One time step in the system
Raw Sensory Data Processed Sensory Data Local and optimised policies Fused Policy
30Hz-1kHz 15-30Hz 30kHz 5-10Hz 30 kHz
Robot Controlled at 1kHz
System-Level Evaluation
System-Level Evaluation
Sense-Plan-Act
System-Level Evaluation
Sense-Plan-Act Feedback Control
System-Level Evaluation
Sense-Plan-Act Feedback Control Full System
System-Level Evaluation
Static Pick and Place Dynamic Pick and Place Dynamic Grasping Dynamic Pointing
System-Level Evaluation
Static Pick and Place Dynamic Pick and Place Dynamic Grasping Dynamic Pointing
Is Motion Planning Overrated?Jeannette Bohg - Interactive Perception and Robot Learning Lab - Stanford
tl;dr; It depends
Environment Complexity
Success of Feedback Control
Environment Complexity
Success of Feedback Control
Environment Complexity
Success of Feedback Control
Environment Complexity
Real-Time Perception meets Reactive Motion Generation. Kappler et al. ICRA’18 + RAL.
Success of Feedback Control
Environment Complexity
Real-Time Perception meets Reactive Motion Generation. Kappler et al. ICRA’18 + RAL.
QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation. Kalashnikov et al. To appear at CORL ’18.
Success of Feedback Control
Environment Complexity
Real-Time Perception meets Reactive Motion Generation. Kappler et al. ICRA’18 + RAL.
QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation. Kalashnikov et al. To appear at CORL ’18.
Motion Planning
Success of Feedback Control
Environment Complexity
Real-Time Perception meets Reactive Motion Generation. Kappler et al. ICRA’18 + RAL.
QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation. Kalashnikov et al. To appear at CORL ’18.
Motion PlanningReactive
Take-Home from Systems Paper
Take-Home from Systems Paper
Higher complexity - More planning
Take-Home from Systems Paper
Higher complexity - More planning
More uncertainty and dynamics - Online Re-Planning
Controlling through contact
Object
Finger
Goal
Controlling through contact
Controlling through contact
Controlling through contact
Controlling through contact
Controlling through contact
Controlling through contact
Better ModelsBetter Feedback
Predictive Model
Model
Predictive Model
Model
Sensory Observations
Predictive Model
Model
Sensory Observations
Action
Predictive Model
Model
Sensory Observations
Action
Predictive Model
Model
Sensory Observations
Action
Predicted Effect
Predictive Model
Model
Sensory Observations
Action
Predicted Effect
Model-Predictive Control
Predictive Model
Predicting physical effect
Goal
Predicting physical effect
Goal
Predicting physical effect
Goal
Keep Predicting
Goal
Keep Predicting
Goal
Our Hypothesis
Our HypothesisPhysics Models
Our HypothesisPhysics Models +
Our HypothesisPhysics Models + Learning
Our HypothesisPhysics Models + Learning
= Generalization
Hybrid Model
Physics-based Model
Action
Predicted EffectParameters
Learned Model
Sensory Observations
Hybrid Model
Physics-based Model
Action
Predicted EffectParameters
Learned Model
Sensory Observations End-to-EndLoss on Effect
Hybrid Model
Physics-based Model
Action
Predicted EffectParameters
Learned Model
Sensory Observations End-to-EndLoss on Effect
Extrapolation
Testing Hypothesis on a Case Study
More than a Million Ways to Be Pushed. A High-Fidelity Experimental Dataset of Planar Pushing. Yu et al. IROS 2016.
Testing Hypothesis on a Case Study
More than a Million Ways to Be Pushed. A High-Fidelity Experimental Dataset of Planar Pushing. Yu et al. IROS 2016.
Testing Hypothesis on a Case Study
More than a Million Ways to Be Pushed. A High-Fidelity Experimental Dataset of Planar Pushing. Yu et al. IROS 2016.
x
x
x
x
K. M. Lynch, H. Maekawa, and K. Tanie, “Manipulation and active sensing by pushing using tactile feedback,” in Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, vol. 1, Jul 1992, pp. 416–421
Analytic Model
x
x
x
x
K. M. Lynch, H. Maekawa, and K. Tanie, “Manipulation and active sensing by pushing using tactile feedback,” in Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, vol. 1, Jul 1992, pp. 416–421
Analytic Model
x
x
x
x
K. M. Lynch, H. Maekawa, and K. Tanie, “Manipulation and active sensing by pushing using tactile feedback,” in Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, vol. 1, Jul 1992, pp. 416–421
1. Stage
Analytic Model
x
x
x
x
K. M. Lynch, H. Maekawa, and K. Tanie, “Manipulation and active sensing by pushing using tactile feedback,” in Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, vol. 1, Jul 1992, pp. 416–421
1. Stage
Analytic Model
x
x
x
x
K. M. Lynch, H. Maekawa, and K. Tanie, “Manipulation and active sensing by pushing using tactile feedback,” in Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, vol. 1, Jul 1992, pp. 416–421
1. Stage
Analytic Model
x
x
x
x
K. M. Lynch, H. Maekawa, and K. Tanie, “Manipulation and active sensing by pushing using tactile feedback,” in Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, vol. 1, Jul 1992, pp. 416–421
2. Stage
1. Stage
Analytic Model
Learned Model
Action
Predicted Effect
Sensory Observations
Neural Network only
Compared Architectures
Learned Model
Action
Predicted Effect
Sensory Observations
Neural Network only Hybrid Model
Physics-based Model
ActionPredicted
Effect
Sensory Observations
Parameters
Learned Model
Compared Architectures
Learned Model
Action
Predicted Effect
Sensory Observations
Neural Network only Hybrid Model
Physics-based Model
ActionPredicted
Effect
Sensory Observations
Parameters
Learned Model
Compared ArchitecturesError Model
Physics-based Model
ActionPredicted
Effect
Sensory Observations
Parameters
Learned Model
Learned Error
+
Learned Model
Action
Predicted Effect
Sensory Observations
Neural Network only Hybrid Model
Physics-based Model
ActionPredicted
Effect
Sensory Observations
Parameters
Learned Model
Compared ArchitecturesRaw Sensory Observations
Error Model
Physics-based Model
ActionPredicted
Effect
Sensory Observations
Parameters
Learned Model
Learned Error
+
Learned Model
Action
Predicted Effect
Sensory Observations
Neural Network only Hybrid Model
Physics-based Model
ActionPredicted
Effect
Sensory Observations
Parameters
Learned Model
Training: End-to-End Loss: Error between Predicted and Ground Truth Effect
Compared ArchitecturesRaw Sensory Observations
Error Model
Physics-based Model
ActionPredicted
Effect
Sensory Observations
Parameters
Learned Model
Learned Error
+
Testing Data Efficiency
Testing Data Efficiency
Testing Data Efficiency
Testing Data Efficiency
Testing Generalization
Alina Kloss et al, “Combining learned and analytical models for predicting action effects,” Submitted. 2018. Pre-print on arXiv.
Testing Generalization
New Pushing Angles & Contact Points
Training Testing
Interpolation
Alina Kloss et al, “Combining learned and analytical models for predicting action effects,” Submitted. 2018. Pre-print on arXiv.
Testing Generalization
New Pushing Angles & Contact Points
Training Testing
New Push Velocities
Training Testing
Interpolation Extrapolation
Alina Kloss et al, “Combining learned and analytical models for predicting action effects,” Submitted. 2018. Pre-print on arXiv.
Testing Generalization
New Pushing Angles & Contact Points
Training Testing
New Push Velocities
Training Testing
New Object Shapes
Training Testing
Interpolation Extrapolation
Alina Kloss et al, “Combining learned and analytical models for predicting action effects,” Submitted. 2018. Pre-print on arXiv.
Training Testing
Generalization to new push velocities
Training Testing
Generalization to new push velocities
Extrapolation
Training Testing
Generalization to new push velocities
Extrapolation
Training Testing
Generalization to new push velocities
Training Testing
Generalization to new push velocities
Extrapolation
Training Testing
Generalization to new push velocities
Training Testing
Generalization to new push velocities
Extrapolation
Training Testing
Generalization to new push velocities
Training Testing
Generalization to new push velocities
Extrapolation
Don’t throw away structure
Learn to extract given state representation from raw data
Physics-based Model
Action
Predicted EffectParameters
Learned Model
Sensory Observations End-to-EndLoss on Effect
A Concrete Suggestion
Interpretability
Interpretability Real and Predicted Box Position after 200 identical pushes
Representation interpretable
Compensation for Errors in Analytical Model Wrong Friction Parameters of Analytical Model
Future Direction
Physics-based Model
Action
Predicted EffectParameters
Learned Model
Sequence of Sensory Observation
Future Direction
Physics-based Model
Action
Predicted EffectParameters
Learned Model
Sequence of Sensory Observation
Multistep Prediction
Future Direction
Physics-based Model
Action
Predicted EffectParameters
Learned Model
Sequence of Sensory Observation
Multistep Prediction
Backpropagation w. r. t. control
Models will never be perfect
Differentiable Recursive Filtering
Jonschkowski et al. RSS’18
Haarnoja et al. NIPS’16
Karkus et al. arXiv’18
Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors. Jonschkowski et al. RSS’18
Learning Heteroscedastic Noise
On Learning Heteroscedastic Noise Models within Differentiable Filtering. Kloss and Bohg. Submitted to ICLR’19. Picture adapted from: Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors. Jonschkowski et al. RSS’18
Learning Heteroscedastic Noise
On Learning Heteroscedastic Noise Models within Differentiable Filtering. Kloss and Bohg. Submitted to ICLR’19. Picture adapted from: Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors. Jonschkowski et al. RSS’18
Observation Noise
Learning Heteroscedastic Noise
On Learning Heteroscedastic Noise Models within Differentiable Filtering. Kloss and Bohg. Submitted to ICLR’19. Picture adapted from: Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors. Jonschkowski et al. RSS’18
Observation Noise
Process Noise
Learning Heteroscedastic Noise
On Learning Heteroscedastic Noise Models within Differentiable Filtering. Kloss and Bohg. Submitted to ICLR’19. Picture adapted from: Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors. Jonschkowski et al. RSS’18
Observation Noise
Process Noise
Learning Heteroscedastic Noise
On Learning Heteroscedastic Noise Models within Differentiable Filtering. Kloss and Bohg. Submitted to ICLR’19. Picture adapted from: Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors. Jonschkowski et al. RSS’18
Observation Noise
Process Noise
Four Non-Linear Filters
Four Non-Linear Filters
Extended Kalman Filter
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Process Model Observation Model
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Process Model Observation Model
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Process Model Observation Model
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Process Model Observation Model
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Process Model Observation Model
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Process Model Observation Model
Four Non-Linear Filters
Extended Kalman Filter
Unscented Kalman Filter (UKF)
Monte-Carlo UKF
Particle Filter
Process Model Observation Model
Loss Function
Loss Function
Ground Truth State Sequence
Loss Function
Ground Truth State Sequence
Sequence of State Estimates (Mean and Covariance)
Loss Function
Ground Truth State Sequence
Sequence of State Estimates (Mean and Covariance)
Network Weights
Loss Function
Ground Truth State Sequence
Sequence of State Estimates (Mean and Covariance)
Network Weights
Negative log-likelihood of true state given believe
Loss Function
Ground Truth State Sequence
Sequence of State Estimates (Mean and Covariance)
Network Weights
Negative log-likelihood of true state given believe Estimation Error
Loss Function
Ground Truth State Sequence
Sequence of State Estimates (Mean and Covariance)
Network Weights
Negative log-likelihood of true state given believe Estimation Error Regularization
Two Application
Kitti Visual Odometry Task Planar Pushing Task
Visual OdometryLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises Mixed
Visual OdometryLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises Mixed
Visual OdometryLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises Mixed
Visual OdometryLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises Mixed
Visual OdometryLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises Mixed
Planar PushingLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises MixedWell-Tuned
NoiseBadly-Tuned
Noise
Planar PushingLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises MixedWell-Tuned
NoiseBadly-Tuned
Noise
Planar PushingLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises MixedWell-Tuned
NoiseBadly-Tuned
Noise
Planar PushingLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises MixedWell-Tuned
NoiseBadly-Tuned
Noise
Planar PushingLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises MixedWell-Tuned
NoiseBadly-Tuned
Noise
Planar PushingLearned Constant Observation Noise
Learned Hsc. Process Noise
Learned Constant Noises MixedWell-Tuned
NoiseBadly-Tuned
Noise
Take Home
Take Home
Differentiable EKF most accurate and robust to inaccurate noise models
Take Home
Differentiable EKF most accurate and robust to inaccurate noise models
Particle filter gains most from heteroscedastic noise model.
Take Home
Differentiable EKF most accurate and robust to inaccurate noise models
Particle filter gains most from heteroscedastic noise model.
UKF works best for more complex dynamics models.
Take Home
Differentiable EKF most accurate and robust to inaccurate noise models
Particle filter gains most from heteroscedastic noise model.
UKF works best for more complex dynamics models.
Take Home
Differentiable EKF most accurate and robust to inaccurate noise models
Particle filter gains most from heteroscedastic noise model.
UKF works best for more complex dynamics models.
Uncertainty bounds for long-term predictions.
Conclusions
Conclusions
Environment Complexity
ConclusionsSuccess of Feedback Control
Environment Complexity
ConclusionsSuccess of Feedback Control
Environment Complexity
ConclusionsSuccess of Feedback Control
Environment Complexity
Motion Planning
ConclusionsSuccess of Feedback Control
Environment Complexity
Motion PlanningReactive
ConclusionsSuccess of Feedback Control
Environment Complexity
Motion PlanningReactive
Better Predictive Models Better Feedback
O P
Thank you for your Attention!
IPRL @ Stanford AMD @ MPI, CLMC @ USCiprl.stanford.edu https://am.is.tuebingen.mpg.de/