Michał Martyniak

My agent is able to navigate from A to B with respect to human input, like: turn left, right, keep lane, go straight (GPS-like navigation). The only input given for the planning module (neural network) is: 224x224 RGB image + command. Output is a ~15m trajectory ahead of agent's car. Most important thing to note is that predicted trajectory does not says "how to drive" but "where is the legal path that should be followed". Besides planning, agent can detect and react to traffic lights (simple perception module made with TensorFlow Object Detection API). Proposed method of Reference Trajectory Approximation allows for better visual understanding how agent percives the environment in the long run. Although the internals of this system are not that much compilcated as it may seem (in comparison with real-world systems), but it took a very long time to do research, gather data, implement and evaluate. More videos + extras (MLinPL): link Feel free to contact me directly if we have common research interests or my work just caught your eye 😉.

We have transferred the data of real-world car traffic to the simulation. The released scenarios cover maneuvers like overtaking on US freeways or exiting roundabouts. Datasets that were used:

• openDD (roundabouts)
• NGSIM (US freeways)

Our research resulted not only in creating the benchamark scenarios but also evaluating different RL policies (reward schemes) and observation types:

• bird's-eye view with ablations
  • front visibility only
  • front+read visibility, without centerline
  • fron+read visibility + framestack (memory)
• RGB camera
• RGB camera + LIDAR

Experimentally, we find that bird's-eye view and dense rewards combined with a penalty for a failure to complete an episode generalizes best to validation scenarios.

My role in the project:

• Main author of the bird-eye's visualization library on PyPI (Github repo)
• Transfer of real-world maps for CARLA (UE4) using RoadRunner
• Co-author of an article accepted to the Autonomous Driving Workshop at NeurIPS 2020 (arxiv version)

Extensive description can be found in this blog post.
Credits: Błażej Osiński, Piotr Miłoś, Adam Jakubowski, Paweł Zięcina, Michał Martyniak, Christopher Galias, Antonia Breuer, Silviu Homoceanu, Henryk MichalewskiBłażej Osiński, Piotr Miłoś, Adam Jakubowski, Paweł Zięcina, Krzysztof Galias, Henryk Michalewski.

TensorHive originated from my university which wanted to solve the problem of conflicts/time collisions between users after purchasing the NVIDIA DGX STATION. Slurm, Kubernetes or other queueing systems would be an overkill (or an unnecessary layer of complexity for students), thus we so came up with something more straightforward.

TensorHive functionalities are as follows:

• Users can use calendar view to make a reservation for a particular set GPUs
• Reserved resources will be protected from running other user's processes
• If reservation gets violeted, the intruder (+admin) will be notified
• It can monitor node resources like: GPU, memory load, etc.

The project has started within GradientPG (Student Research Group)