Robot Path Planning with Dynamic Obstacle Prediction using SGANs

Team # 5

Affan Bin Usman | Chinmay Bhale | Pooja Bharathi Oguri | Vishnu Pisharam

PROBLEM STATEMENT

Problem

• Robot motion planning in dynamic environments is a challenging problem
• Requires the robot to generate
• Safe & efficient paths
• Avoiding collisions with moving obstacles
• Starship food delivery robots
• Stops if people come in front of the robot

Motivation

• Traditional motion planning algorithms do not handle dynamic obstacles
• Leads to unsafe and inefficient navigation
• Novel approach that can integrate
• Dynamic obstacle predictions
• Traditional motion planning algorithms

2

SOLUTION

• Dynamic obstacle prediction using SGANs
• RRT* algorithm + Spline Interpolation for path planning
• Motion control with Non-linear MPC
• Goal
• Safety of pedestrians
• Efficiency of autonomous robots in dynamic environments

3

BASELINE

• GANs to overcome the difficulties in
• Approximating intractable probabilistic computation
• Behavioral inference

• Improve performances of social robots & self-driving cars
• Predict motion behavior of pedestrians

* Paper: Social GAN - Socially Acceptable Trajectories with Generative Adversarial Networks

• Input
• Vector of previous trajectories
• Output
• Pooled vector for each individual.

• Structure
• Generator: RNN Encoder-Decoder
• Discriminator: RNN-based encoder

4

BASELINE

• Inputs (decoder)
• Histories
• Pooled vector
• Output (decoder)
• Future trajectory

* Paper: Social GAN - Socially Acceptable Trajectories with Generative Adversarial Networks

5

BASELINE

• Inputs (decoder)
• Histories
• Pooled vector
• Output (decoder)
• Future trajectory

* Paper: Social GAN - Socially Acceptable Trajectories with Generative Adversarial Networks

• Discriminator determines if the generated trajectories are real or fake
• Training dataset
• 5 sets of publicly available datasets (ETH, Hotel, Univ, Zara1, Zara2)
• Leave-one-out approach
• Error Calculation - Evaluating accuracy of the generated trajectories
• Average Displacement Error (ADE)
• Final Displacement

6

* Paper: Social GAN - Socially Acceptable Trajectories with Generative Adversarial Networks

Fig 1: Error Plot

Fig 2: Error for Datasets & Models

Table 1

7

What to do with these Predictions?

8

Predictions of humans

Model Dynamic Obstacles

Robot’s Sensory Input

Path Planning with Obstacle Avoidance

Reach at Desired Location

WORKFLOW

Previous 8 Steps

Predicted Future

Trajectories

Sends List

Finds Path

Smoothes Out Path

Next 8 Steps of Robot

9

Calculate Obstacle List

SGAN

Simulation

(Humans & Robot)

RRT* / Custom RRT*

Spline Interpolation

NMPC

PATH PLANNING

Fig 1: Path with RRT*

Fig 2: Path with modified RRT*

10

MOTION PLANNING

Obstacle Avoidance with NMPC

11

SIMULATION STRUCTURE

Front End

Back End

12

Person.js

SGAN

Sim.js

Robot.js

Main.py

RRT*

Evaluate.py

NMPC

Docker

Robot Class

Back End

Front End

Person.js

SGAN

Sim.js

Robot.js

Main.py

RRT*

Evaluate.py

NMPC

Docker

Robot Class

PEOPLE SIMULATION

14

15

Robot within threshold distance of 1.5m considered as collision (Due to human safety consideration)

FUTURE WORK

• 2D → 3D
• Testing & implementation in real-world environment
• Evaluate with respect to Extended Social force model.
• Comparison against different path planning approaches

16

Thank you!

17