Abstract
This proposed larger-scale effort aims to re-define and demonstrate the vision of full autonomy to one of safe autonomy, where a learning-enabled
system is coupled with the foundations of cyber-physical systems to endow the system with an explicit awareness of both its capabilities and limitations. In turn, the system realizes when it is in or near a zone where its safety cannot be assured, and thereby transitions to a safe fallback state. A multi-pronged approach is adopted to achieve safe autonomy: (a) creating contextual awareness of the operating conditions to modify learning- and logic-based behaviors to reflect the operational context; (b) determining the location and orientation of the AV in absolute and relative coordinate frames to serve the needs of different tasks reliably and scalably; (c) defining and enforcing both static and dynamic guards for safe real-time actuation; (d) developing a powerful co-simulation framework to safely and efficiently test system performance under a range of clear and adverse operating conditions; and (e) validating and demonstrating our methodology on CMU’s Cadillac CT6 autonomous vehicle. The effort will also showcase physical demonstrations of vehicle capabilities to researchers and visiting dignitaries.
Recent advances in machine learning (ML) have been significant, and the application potential for ML seems limitless. However, using ML in its current form inevitably generates a non-zero amount of false positives and negatives, which in a safety-critical system can potentially be disastrous, causing damage to life and/or property. At the same time, the judicious use mathematical foundations, scientific principles and engineering ingenuity has led to the creation of large-scale but practical safety-critical systems such as aviation, nuclear power plants, electric grids and medical devices. In this effort, we build on the conjecture that learning-enabled systems must necessarily be guided and fenced by logical, explainable and analyzable safeguards. Specifically, we propose to apply our methodology to the domain of connected and autonomous vehicles which must address a very long tail of known and unknown scenarios.
Description
Timeline
Strategic Description / RD&T
Autonomous vehicles have held the promise of ”Vision Zero”, where there are zero fatalities from automotive crashes. This holy grail has been intended to be accomplished by CPS and AI capabilities that never get distracted, tired or drunk, thereby circumventing the root cause of human error in most automotive fatalities. If vehicles can drive themselves, members of society who are differently abled, legally blind or otherwise unable to drive themselves can experience a much better standard of living by not being dependent on others for their transportation needs. Furthermore, if vehicles drive themselves, commute times for those who have to be responsible for driving to and from work can enhance their productivity and experience the benefits of having a virtual chauffeur. As we now know, however, these promises have proved to be both elusive and expensive to accomplish. Hence, there is a strong need to re-think what autonomy and safety-critical operations mean. This proposed effort aims to lay strong foundations and guiding principles that over time will enable safe operations and adoption of autonomous vehicles.
Deployment Plan
Q1: Development of a Safety Assurance Framework
Q2: Testing in Simulation Environment
Q3: Validation in Workzone Environment
Q4: Demonstration in Workzone Environment
Expected Outcomes/Impacts
While driver-assist and other safety features are proving helpful and increasingly desirable, the full potential of AVs has not been fully realized yet due to several factors: (a) these are inherently safety-critical systems; (b) the complexity of driving is extremely high, with a very long tail of known and unknown ”edge cases” that must be dealt with; (c) societal acceptance will follow trustworthiness of AVs; and (d) with the technology still evolving, regulatory frameworks for AVs are not in place yet. AVs, by necessity caused by the lack of algorithmic solutions in many cases, rely heavily on machine learning for perception-related tasks like object detection, object classification at the level of an image pixel, and the reliable detection of traffic signal status. Unfortunately, the fact that learning-based systems produce a non-zero number of false positives and false negatives runs squarely against the safety-critical aspect of driving. In other words, vehicles with higher-levels of autonomy are learning-enabled and must also satisfy stringent safety requirements necessitated by the potential loss of life and/or property when failures occur. Correspondingly, AVs have been considered to be a grand challenge for both the domains of cyber-physical systems and artificial intelligence. With the USDOT, FHWA and ARPA-I focusing (or expected to focus) on autonomy and connected autonomy, the realization of revolutionizing transportation is still many years away. Expectations in terms of total autonomy, the use of only vehicle-based sensors, and the deployment timelines need to be moderated, even as agencies and communities are shown that autonomous vehicles can be (much) safer than human drivers. The road may be longer than many have promised, but we strongly believe that a more practical and acceptable mode of operation can be realized in a few years. This proposed effort builds on our experience of nearly 20 years to move the state of the art in that direction and timeframe.
Expected Outputs
Autonomous vehicles have held the promise of ”Vision Zero”, where there are zero fatalities from automotive crashes. This holy grail has been intended to be accomplished by CPS and AI capabilities that never get distracted, tired or drunk, thereby circumventing the root cause of human error in most automotive fatalities. If vehicles can drive themselves, members of society who are differently abled, legally blind or otherwise unable to drive themselves can experience a much better standard of living by not being dependent on others for their transportation needs. Furthermore, if vehicles drive themselves, commute times for those who have to be responsible for driving to and from work can enhance their productivity and experience the benefits of having a virtual chauffeur. As we now know, however, these promises have proved to be both elusive and expensive to accomplish. Hence, there is a strong need to rethink what autonomy and safety-critical operations mean. This proposed effort aims to lay strong foundations and guiding principles that over time will enable safe operations and adoption of autonomous vehicles.
Most importantly, concepts must be realized on a functional but highly complex platform. Having clearance from both state and city government agencies to test on public roads, a "license" that we know of no other academic group in the entire nation as having one, we are uniquely positioned to demonstrate our capabilities to Safet21 sponsors, the research and user communities, as well as visitors to Safety21 and CMU.
TRID
Individuals Involved
| Email |
Name |
Affiliation |
Role |
Position |
| rajkumar@cmu.edu |
Rajkumar, Raj |
CMU |
PI |
Other |
Budget
Amount of UTC Funds Awarded
$285000.00
Total Project Budget (from all funding sources)
$650000.00
Documents
Match Sources
No match sources!
Partners
| Name |
Type |
| RIDC |
Deployment Partner Deployment Partner |
