도요타 AI Center for AI Research (재미있는 연구 주제 들)

도요타 AI Center for AI Research (재미있는 연구 주제 들)

Human-Autonomous Vehicle Systems
- Reactive Collision Avoidance / Interaction Aware Planning / I-POMDP
- Investigate interactions across a range of timescales
- Long Timescales ; Interactive Partially Observable
- Markov Deicsion Procss (I-POMDP) for Navigation
- Intermediate Timescales : Interaction-aware trajectory planning using conditional variational auto encoders and game theoretic planning
- Short Timescales : Reactive collision avoidance using buffered Voronoi cells and reachability analysis
- Combine algorithms into integrated autonomy stack 
- Track experiments with full scale autonomous vechicles
- Engagement with TRI fro tech transfer  and infusion

Closing the Visual-Motor Loop with Deep Reinforcement Learning
- Navigation in multi-agent settings : We will explore end-to-end sensorimotor learning in a real-world task.
- Boostrapping with Supervised learning : We will move from the simpler task of navigation with pedestrian avoidance to more complex tasks such as driving a simulated autonomous vehicle.
- Refinement with Active Reinforcement Learning: We will relax the fully supervised assumption by supervised initialization and fine-tuning the entire system using reinforcement learning algorithms.

Design new knowledge representations enabling pervasive ambient intelligent systems, such as a smart house or smart car cabin
- Representation of knowledge about humans and their environments in homes, offices and vehicles.
- Acquisition of this knowledge from sensor data and its organization / indexing for efficient retrieval.
- Ability to transport acquired knowledge to new settings in a highly customizable and specific fashion, especially in crossmodality (e.g. from 3D scan to image) and cross-domain (e.g. from home to vehicle) cases.
- The ability to scale to big data collections recordings multiple diverse environments and human actions in them.

Task-Driven Visual Perception
- Integrating visual perception as an adaptable and evolving module within a larger robotic framework.
- Developing a perception model that can extend its skill set to specifically support the downstream job of a robot
- Gearing the perception model with:
    - Generatlization mechanisms for solving novel perceptual task faced at the execution time, driven by the downstream robotic goal.
    - Adaption mechanisms to improve and specialize during execution

Joint Contextual Forecasting of Vehicle Behaviors, Actions, and Trajectories
* Dataset Collection: Using open datasets such as Berkeley Deep Drive (BDD), and exploring simulators such as Carla for data collection for benchmarking.
* Forecasting vehicle trajectories, actions, and behaviors: Implementing several baseline methods, along with novel methods (based on SSL and IL) that independently forecast each concept.
* Joint Learning and Understanding: Developing methods to jointly learn the above vehicle dynamics.
 
Uncertainty on Uncertainty, Robustness, and Simulation
1. Robustness against “known unknowns” (rare events already present in the training data)
    1. Convergence rates for robust risk
    2. Importance sampling for robust risk evaluation
    3. Calibration and certification of risk
    4. Optimization and design
2. Robustness against “unknown unknowns” (rare events not present in training data)
    1. Introducing additional noise in training data
    2. Risk-sensitive importance sampling
    3. Robustness against actions of others

Understanding Driver State in Laboratory and Naturalistic Environments
- Develop testbeds for collecting driver information in lab simulator and on-road
- Collect human-centric multimodal data from semi-controlled, state induced experiments
- Create a corpus of driver state data ready for ML community
- Begin modeling and testing future driving interactions based on driver state information

References
[1] https://aicenter.stanford.edu/research