Technological Perspective on Human‑aware Navigation Algorithms
Perspectives on Human‑aware Navigation
International Conference on Social Robotics
Madrid, 26 November 2019
Motivation:
Facts:
- Increasing number of factors to consider
- Increasing code complexity
Questions:
- Wouldn't ML be easier and more cost-efficient?
- Do humans consider factors unknowingly? (people might be unaware of other factors)
Not any ML algorithm is enough!
- Heterogeneous: many factors
- A variable number of people
- A variable environment, including objects
- A variable number of interactions
- A variable number of types of interactions
- Indeterminately complex & structured relationships
ML: Handcrafted features?
- Variables considered: we can let ML learn to select
- How to come up with features for the previous data?
- A variable number of people
- A variable environment, including objects
- A variable number of interactions
- A variable number of types of interactions
- Indeterminately complex & structured relationships
ML: End-to-end learning?
- E2E doesn't have some the mentioned problems
- They have other problems
- Huge amounts of data
- Computation resources(especially but not only for training)
- How to deal with structured data?
Graph Neural Networks
Function approximators which take graphs as input
Graph Neural Networks for
Human-aware Navigation
- Improve accuracy of other ML algorithms
- Improve scalability(how can we increase the number of variables to consider?)
Graph Neural Networks for
Human-aware Navigation
- Model proxemics / inconvenience
- Predict people's movements
- Control robot's movements
- Detect & predict behaviours/events
SocNav1 Tool
- The idea of the dataset
- Screenshot
- Labelled scenarios: 9280
- Limitations
- Humans are static
- One type of interaction
- We are told "how people think they would feel"
SocNav1 Results
SocNav1 Results
Experiment with GNNs & SocNav1 dataset
- w-r:
world→room - r-R:
room→robot - h-R:
human→robot - o-R:
object→robot - h-h:
human→human - h-o:
human→object
Social Navigation Graph Neural Network (SNGNN)
Social Navigation Graph Neural Network (SNGNN)
Results: Personal Spaces & Interactions
- Personal spaces
- Interactions
Results: Distance between two interacting people
Results: Impact of walls
Results: Impact of incrementing the number of people in a room
Results: Angle when approaching interacting people
Other works using GNNs
Relational Graph Learning for Crowd Navigation
C. Chen, S. Hu, P. Nikdel, G. Mori, M. Savva
- Relational Graph Learning for Crowd Navigation. Changan Chen, Sha Hu, Payam Nikdel, Greg Mori, Manolis Savva. arXiv preprint arXiv:1909.13165. 2019.
Learning Human-Object Interactions by Graph Parsing Neural Networks
S. Qi, W. Wang, B. Jia, J. Shen, S.-C. Zhu
Conclusions
- New dataset
- We will be hearing more about GNNs(Deep Mind, Google Brain)
References
Graph Neural Networks:
- Semi-Supervised Classification with Graph Convolutional Networks. T.N. Kipf, M. Welling. arXiv preprint arXiv:1609.02907. 2017.
- Relational inductive biases, deep learning, and graph networks. P.W. Battaglia et al.. arXiv preprint arXiv:1806.01261. 2018.
Applications:
- Graph Neural Networks for Human-aware Social Navigation. L.J. Manso, R.R. Jorvekar, D.R. Faria, P. Bustos, P. Bachiller. arXiv preprint arXiv:1909.09003. 2019.
- SocNav1: A Dataset to Benchmark and Learn Social Navigation Conventions. L.J. Manso, P. Nunez, L.V. Calderita, D.R. Faria, P. Bachiller. arXiv preprint arXiv:1909.02993. 2019.
- Learning Human-Object Interactions by Graph Parsing Neural Networks.S. Qi, W. Wang, B. Jia, J. Shen, S.-C. Zhu. arXiv:1808.07962. 2018.
- Relational Graph Learning for Crowd Navigation. C. Chen, S. Hu, P. Nikdel, G. Mori, M. Savva. arXiv preprint arXiv:1909.13165. 2019.