Utilising Semantically Meaningful Predictions in Motion Planning
Lead Research Organisation:
University of Edinburgh
Department Name: Sch of Informatics
Abstract
Automated decision making is becoming widely used in applications ranging from automated driving, construction and video surveillance to even health-care. We are still far away from being able to fully replicate the fluidity and competence of human behaviour in these same settings. A key stumbling block is around interactions with a dynamic environment, e.g., when the environment includes other decision making agents.
In this work, we are interested in building machines that can move intelligently in dynamic environments by better utilising semantic information, such as room layouts and how this affects others' motion.
This requires a number of technical problems to be solved. The most important one is the need to predict others' future actions, i.e., intent. These predictions must be conditioned on what is observed in the environment, e.g., that a person may deviate from a straight-line path due to a visible obstacle. The second problem is to incorporate these predictions into a motion planning algorithm, to make decisions for one's own. We approach this problem through a combination of neural network based predictive models and reinforcement learning that utilises such models. This also enables further possibilities, such as to categorise types of motion and to detect anomalies.
The primary contributions of our work include:
- Improved predictive models that utilise environmental features and agent motion to predict future behaviour;
- Motion planning techniques that utilise such predictions, such as new model-based reinforcement learning algorithms;
- Behaviour classification and anomaly detection from video input.
In this work, we are interested in building machines that can move intelligently in dynamic environments by better utilising semantic information, such as room layouts and how this affects others' motion.
This requires a number of technical problems to be solved. The most important one is the need to predict others' future actions, i.e., intent. These predictions must be conditioned on what is observed in the environment, e.g., that a person may deviate from a straight-line path due to a visible obstacle. The second problem is to incorporate these predictions into a motion planning algorithm, to make decisions for one's own. We approach this problem through a combination of neural network based predictive models and reinforcement learning that utilises such models. This also enables further possibilities, such as to categorise types of motion and to detect anomalies.
The primary contributions of our work include:
- Improved predictive models that utilise environmental features and agent motion to predict future behaviour;
- Motion planning techniques that utilise such predictions, such as new model-based reinforcement learning algorithms;
- Behaviour classification and anomaly detection from video input.
People |
ORCID iD |
| Todor Davchev (Student) |
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/N509152/1 | 01/10/2015 | 31/03/2021 | |||
| 1827837 | Studentship | EP/N509152/1 | 01/09/2016 | 28/02/2021 | Todor Davchev |
| Description | AI solutions in general can benefit from the introduction of inductive biases. For example, the context in which moving agents interact plays a significant role in the generation of their motion. Through the research funded on this grant we have discovered that utilising a learned model (i.e. an inductive bias) of the environment can facilitate long-term predictions, even if that model was learned in a modular manner. Further, we observed that given an accurate abstraction of the static surrounding, the proposed solution can generalise to unseen environments better than state-of-the-art alternatives. We highlight the model's prediction capability using a pedestrian tracking problem and a tabletop manipulation task where trajectories are generated conditioned on a plan learned from observations of another agent. Further, we have shown that a similar approach can be taken to achieve stronger adversarial robustness across tasks (such as classification of objects) even in scenarios with little data. We go on to show that such inductive biases are helpful in the field of Robotics too in the context of manufacturing, where roboticised solutions can be made more robust to environmental changes. |
| Exploitation Route | This project is established through an iCASE award with Thales. The expectation is that this work will be applicable to autonomous vehicles of various kinds - immediately for road vehicles that need pedestrian predictions as part of the environment models, but also for other, i.e. maritime vehicles, that also operate in dynamic environments. We are currently exploring options around internships for skills transfer and Thales has raised a potential interest in having me as intern where I could apply my methods to their data. More broadly, such predictions are useful in the toolchain for human-robot interactions. |
| Sectors | Aerospace, Defence and Marine,Digital/Communication/Information Technologies (including Software),Financial Services, and Management Consultancy,Healthcare,Manufacturing, including Industrial Biotechology,Transport,Other |
| URL | https://arxiv.org/abs/1911.13044 |