Periodic Reporting for period 1 - NeuroMotive (NeuroRobotics Platform Application to Traction Control for Automotive Industry)
Período documentado: 2018-07-01 hasta 2019-12-31
NeuroMotive provides a tailor-made service for automotive that exploits the synergy between the repeatable testing conditions afforded by the Hardware in the Loop infrastructure and cutting-edge AI Tools for Machine Learning to support the automated development of high-performance Traction Controllers, and of Dynamical Models for precise emulation of the car’s road behaviour (Numerical Twin for DDT).
The NeuroMotive project strives to achieve these aims by combining the strengths of two powerful tools:
a) The above mentioned Siemens Mobility Test bed (SMTS) which – as an automotive Hardware in the Loop Infrastructure - provides a controlled environment supporting precise, repeatable testing conditions, and automation of data-set generation for Deep-Learning. This test bed is unique in Germany (just one comparable infrastructure exists in industry at Valeo) because each of the four wheels can be addressed and controlled independently. A RODING car is used a demonstrator on top of this testbed to demonstrate the impact of signals set for instance by neural networks on the motion of a car under natural road conditions. The hardware in the Loop automotive setup allows for control-function tuning, validation and performance assessment. The integration of the considered vehicle within the testing setup allows physical emulation of a wide range of road conditions for any car connected.
b) The cutting-edge AI platform, NeuroRobotics Platform (NRP), developed in the HBP FET-Flagship project which supports AI Libraries (e.g. TensorFlow, NNabla), Neural Engines (NEST, Nengo), Virtual Twin functions, a robust suite of model ID and control tools. These tools for AI and Machine Learning allow neural controller development and tuning, Model Identification / Virtual Twin using the NeuroRobotics Platform Tool Suite. The platform directly supports controller synthesis, and machine/deep learning: tuning, adaptation of model parameters and control gains.
Concrete services to the automotive industry and suppliers were identified during the runtime of the project with intensive exchange with relevant stakeholders on the market based on the existing network of the partners.
• The SMTS hardware platform itself: with its almost unique feature of being able to drive each wheel independently. This platform is suited to testing relatively small cars due to the power limitations of the motors.
• The NRP software: which can enhance the hardware offering by allowing extended testing in scenarios that have not been physically encountered and recorded in on-road testing with hardware in the loop simulation testing. In the longer term, the NRP software could be used to develop and tune controllers for cars to
provide enhanced driver assistance and semi-autonomous and autonomous driving.
• The expertise built up within TUM that can be utilised to provide both SMTS related services but also specialised R&D projects based around developing further the capabilities of SMTS enhanced with NRP based software.
Quite a lot of effort has gone into setting up the Siemens testbed at TUM. Obstacles which had to be overcome and could not be foreseen included major adjustments of the building to accommodate the test bed, i.e.:
1. fortification of the concrete floor of the lab (including bottom & material sampling),
2. purchase of a lifter to lift the test cars to the flanges of the test bed,
3. solid wall with a bullet-proof window to protect researchers/users from injury from components flying away due to gravity, together with an industrial door to allow cars to access the robotics lab.
Challenges were posed not only by the implementation itself but also by the need to identify the right suppliers and by full applicability of the tender regulations for each individual task to be subcontracted. Apart from the physical setup, the electronic relaunch of the testbed has to be managed. This included the selection of the right software, the identification of potential suppliers and respecting the tender regulations when subcontracting the implementation.
The project successfully delivered the work related to concept determination. The practical demonstration was not feasible, though, due to the limitations in the function of the tools. These can be expected to be overcome shortly, but unfortunately, it was not possible to do so during the runtime of the project.
While the long term impact of NeuroMotive remains much the same, particularly because the project confirmed the relevance of Reinforcement Learning, the robotics industry is considering with growing interest the technical challenge of the sim-to-real transfer. The shorter-term impact will be tempered by the spinning out of the less ambitious offerings of SMTS testing of small vehicles (capitalising on its near-unique features) coupled with a series of R&D projects aimed at achieving synergistic coupling of the NRP software with the SMTS platform. This will achieve two aims:
• First, building stronger links and alliances with key automotive players both as customers and collaborators;
• Second, it will enable the building of capability that will result in a truly unique offering that was first foreseen in the NeuroMotive project.
Last but not least the project revealed relevance for education and for further R&D projects, with industry, on a national German level, but also on the European level. Leveraging dedicated simulators like AirSim is therefore a particularly attractive option. At any rate, this topic must become an area of focused study for the development of the envisioned services.