#471 Safe decision-making in interactive environments

Principal Investigator
yorie nakahira
Start Date
July 1, 2023
End Date
June 30, 2024
Project Type
Research Advanced
Grant Program
US DOT BIL, Safety21, 2023 - 2028 (4811)
Grant Cycle
Safety21 : 23-24


Autonomous ground vehicles must safely operate in highly interactive environments with human uncertainties. Safe actions depend on context, interactions, and absolute (mathematical measures for safety) may differ from how humans perceive as safe and reliable behaviors. However, producing context-dependent and interaction-aware safe actions is non-trivial due to the following technical challenges. 

Challenge 1: Hardness in the identification of interaction mechanisms. When interaction mechanisms are known or can be learned, many safe learning and control techniques can be employed. However, in many interactive environments, it may be fundamentally difficult to fully identify the mechanisms of the opponent's interaction due to unobserved confounders. 

Challenge 2: Latent risk and latent variables. There exist latent risks (such as occlusions) and unobserved variables (e.g. awareness and intention), and safe actions depend on such factors. For example, pedestrians can decide to enter a crosswalk, but such intent may not be directly observable.  Decisions that ignore such factors may experience unexpected risks. 

Challenge 3: Tensions between long-term safety vs. computation. Accounting for long-term outcomes in risk quantification and control is challenging due to stringent computation vs. time-horizon tradeoffs, particularly for rare events. Myopic safety can be efficiently certified, but ensuring long-term safety may require prohibitive computation.

Although distribution shifts are often approached by finding actions that are robust to changes or avoid changes, much less work explores how to proactively induce desirable opponent behavior changes in interactive environments. For example, whereas humans can slowly squeeze their way through crowded environments, autonomous systems programmed to maintain worst-case distance may not find feasible solutions (e.g., freezing robot problems). Proximity can be safe or risky depending on opponents’ interaction mechanisms, but this is not captured in a simple risk measure of distance. The proposed research aims at realizing such capabilities by accounting for interaction models and anticipating unobserved unknowns in risk quantification and decision-making. Specifically, we propose the following research.  

Thrust 1: We will establish an efficient risk quantification method with theoretical guarantees. We will leverage an integrative view of stochastic systems, MC, PDEs, and Physics-informed neural networks (PINNs) to estimate intervention risk from heterogeneous data and exploit low-dimensional structures. Our prior work has derived four types of long-term safe probabilities as unique solutions to deterministic linear PDEs, which characterize the relation between risk probabilities of different time horizons and initial states. PINNs with these PDE constraints have been shown to be able to infer risk probabilities beyond available data with provable generalization. Here, we will explore such characterization and enable long-term risk to be quantified using shorter-term interaction data. 

Thrust 2: We will develop efficient long-term safety certificates for interactive environments. While treating statistical models as mechanistic models may neglect important latent risk factors, little effort has been made to rigorously differentiate the underlying mechanistic models vs. observed statistical models in the design of safety certificates. Here, we will build upon our prior work on probabilistic invariance to differentiate the two models, control the risk probability using observed statistics, and handle information constraints arising from delayed and rate-limited communication. 


Strategic Description / RD&T
Contribution to the US DOT. 
The proposed research is expected to contribute to US DOT’s strategic goals in safety, equity, and transportation. 
- Safety: The proposed techniques are expected to improve the safety of autonomous vehicles operating in highly interactive situations while anticipating risks from unobserved factors. The research is expected to apply to connected autonomy, where multiple autonomous vehicles must safely navigate crowded environments with pedestrians and human drivers. 
- Equity: The proposed techniques are expected to be implementable in low-cost devices due to a more optimal allocation of onboard and communication resources. Such techniques can be useful for autonomous ground vehicles used for delivery, navigation for the blind, and surveillance, thereby promoting affordable and accessible transportation systems.   
- Transformation: The proposed techniques can be integrated into learning-based techniques to leverage the capability of data-driven methods to handle complex interactions with rigorous risk quantification and assured safety. Transparency to the exposed risk and the safe use of data-driven techniques will help modernize transportation systems. 
These contributions also align with US DOT’s research priorities in assistive and accessible mobility; equity, affordability, and accessibility, as well as traffic safety enforcement. 

Contribution to the Safety21 UTC focus: 
- Connected autonomy: The proposed safety certificates or safe planning and control techniques apply to autonomous vehicles operating under human uncertainties, limited onboard resources, and delayed or rate-limited communication. 
- Physical and digital infrastructure: The proposed risk quantification techniques can be deployed in transportation digital infrastructure to access risk in real-time. 
Deployment Plan
Quarter 1: We will develop the following three test beds. 
Simulated environment for navigation and interaction. We have been using CARLA in past research, and will additionally use SUMMIT to enable high-fidelity test environments for dense urban traffic.
Virtual environment games. A virtual driving platform is currently being developed. Human agents sit in real-sized vehicles to drive in virtual environments or experience autonomous driving.

Quarter 2: We will deploy the proposed technique in multi-agent navigation in diverse scenarios. The deployment will primarily focus on showcasing the proposed algorithms' safety, performance, and robustness. The experiments will consider autonomous ground vehicles navigating a crowded space with occlusions, pedestrians, and other vehicles. In some scenarios, agents must travel through spaces so crowded with other agents that safe actions do not exist for all worst-case uncertainties, and the vehicles must take risks to reach the destination. The learning-based controllers may not always produce safe actions, and the proposed safety certificate is expected to modify such unsafe actions. 

Quarter 3: We will perform experiments with human subjects. The deployment will primarily focus on generating behaviors that are not only safe but for humans to feel safe. Human subjects will be asked to drive in virtual environments or experience autonomous driving. The results will be used to produce safe and human-like behaviors that improve the perceived safety of human users, human drivers, and pedestrians on the algorithms.  
Expected Outcomes/Impacts
Safety: The proposed techniques are expected to navigate crowded, interactive environments with better safety and feasibility. The performance, and robustness of the proposed methods will be evaluated and compared to the existing techniques for autonomous ground vehicles operating in diverse environments. 

Reliability: Be better account for human driving models, the proposed techniques are expected to provide more comfort for humans sitting in automated vehicles. The reliability will be tested in human experiments sitting in real-sized vehicles virtually experiencing automated driving. We will use the experiments to understand 

Durability and cost: The proposed research studies how to produce safe actions using reduced computation and unreliable communication. These efforts will contribute to the durability and cost-reduction of intelligent vehicles and connected autonomy. Through deployment and case studies, we will also study additional design considerations such as optimal allocation of onboard resources, offloading computation offline, and deployment in embedded devices.  
Expected Outputs
We anticipate the following outputs. 

(1) Theory, methodology, and algorithms for risk quantification, control, and learning techniques that jointly realize the following merits.
Merit 1: Produce safe and interaction-aware actions 
Merit 2: Anticipating risks from unobserved factors
Merit 3. Physics-Informed Risk Quantification with provable generalization
Merit 4: Interaction-Aware Safety Certificate without Complete Interaction Mechanisms

(2) Integrated software package, data sets from experiments, and testbed for future research: The code for the proposed techniques, experimental data, and testbed will be made publicly available through publications. We will include detailed descriptions so that they can be useful for research, engineering, and education. These tools will also be disseminated in PI’s classes for students to better understand available techniques for safe interactions in intelligent transportation systems. 

(3) A set of demonstrations that illustrate the capability to handle complex interactions and latent risks. The proposed system is expected to exhibit flexible behaviors such as slowly squeezing forward a way through a congested area, maintaining sufficient distance from agents who do not actively avoid obstacles, following the flow of a crowd (known as a flow-following strategy), reducing speed in anticipation of latent risks (occlusions). 
Our past project (see attached TRID document) has established a basic framework for the design of safety certificates that ensure long-term safety. The proposed work will expand our prior work in the following sense. 
1. We will generalize this framework to account for interactions and behavior models of pedestrians and other drivers.
2. The proposed framework and many safe learning and control techniques rely on accurate risk quantification, which has not been studied in our past project. The proposed research will also develop estimation techniques for risk probabilities, to be integrated into the decision-making pipeline. 
3. We will additionally study whether the proposed methods will be perceived by humans as safe and reliable. 

Individuals Involved

Email Name Affiliation Role Position
maitham@cmu.edu Alsunni, Maitham CMU Other Student - Masters
jbelke@andrew.cmu.edu Belke, Jordan CMU Other Staff - Business Manager
helensloeb@gmail.com Loeb, Helen Jitsik LLC Other Other
yorie@cmu.edu nakahira, yorie CMU PI Faculty - Untenured, Tenure Track


Amount of UTC Funds Awarded
Total Project Budget (from all funding sources)


Type Name Uploaded
Data Management Plan Data_management_plan_diupxZ9.docx Oct. 23, 2023, 8:43 a.m.
Publication PINN.pdf March 29, 2024, 12:29 p.m.
Publication Context-aware LLM-based Safe Control Against Latent Risks March 29, 2024, 6:23 p.m.
Presentation video_presentation.pdf March 29, 2024, 12:29 p.m.
Progress Report 471_Progress_Report_2024-03-31 March 29, 2024, 6:23 p.m.

Match Sources

No match sources!


Name Type
Jitsik LLC Deployment Partner Deployment Partner
City of Pittsburgh, Department of Mobility & Infrastructure Equity Partner Equity Partner