Multi-Objective Multi-Agent Planning for Jointly Discovering and Tracking Mobile Objects
Authors: Hoa Van Nguyen, Hamid Rezatofighi, Ba-Ngu Vo, Damith C. Ranasinghe7227-7235
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments We evaluate the proposed value function using a series of comprehensive synthetic experiments since we can control all of the parameters of the problem, especially with a timevarying number of agents and objects. We compare three planning algorithm formulations: (i) using the single objective value function V1( ) in (4) for tracking. (ii) using a single objective value function based on our new discovery value function V2( ) in (10). (iii) using our proposed multi-objective value function Vmo( ). We use optimal sub-pattern assignment (OSPA) (Schuhmacher, Vo, and Vo 2008) to measure performance. We report OSPA Dist as the main metric to evaluate the overall team performance since it incorporates both tracking and discovery indicators. For further insights into our planning formulations, we also report: (i) OSPA Loc as a localization accuracy measure, (ii) OSPA Card as an object discovery performance measure; and (iii) Search Area Entropy as the average entropy of the occupancy grid to measure the coverage area of the team. For demonstration, a team of quad-copter UAVs flying at different altitudes is considered. The detailed parameter settings are provided in the appendix, while scenario setups are shown in Figure 1. Our experiments considered four different scenarios and two different detection-based sensors subject to noisy measurements. |
| Researcher Affiliation | Academia | 1The University of Adelaide 2Curtin University {hoavan.nguyen, hamid.rezatofighi, damith.ranasinghe}@adelaide.edu.au, ba-ngu.vo@curtin.edu.au |
| Pseudocode | No | The paper refers to using a "greedy search algorithm see the work by (Dames, Tokekar, and Kumar 2017)" but does not provide the pseudocode for it within the paper. |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code or links to a code repository. |
| Open Datasets | No | The paper states: "We evaluate the proposed value function using a series of comprehensive synthetic experiments since we can control all of the parameters of the problem, especially with a timevarying number of agents and objects." It describes the generation of this synthetic data (e.g., constant velocity model, detection probability, sensor noise parameters) in the Appendix but does not provide public access to the generated datasets or use an established public dataset. |
| Dataset Splits | No | The paper describes running "20 MC runs" for each scenario but does not specify explicit train/validation/test dataset splits. The experiments are simulation-based, where data is generated for each run rather than split from a pre-existing dataset. |
| Hardware Specification | No | The paper mentions "quad-copter UAVs" as the mobile agents in the simulation but does not provide any specific details about the computational hardware (e.g., GPU/CPU models, memory) used to run the planning algorithms and simulations. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers needed to replicate the experiments. |
| Experiment Setup | Yes | Appendix Parameter settings for experiments The search areas for the first three scenarios and scenario 4 are 1000 m 1000 m and 2000 m 2000 m, respectively. Each agent is controlled to fly at a fixed and different altitude (i.e., 5 m altitude gap between each agent) to prevent collisions with other team members. The minimum altitude starts at 30 m for the first agent and increases 5 m for each additional agent. Further, all objects are assumed exist on a horizontal ground plane to speed up the numerical experiments by tracking in 2D. Each object state x = (x, l) is uniquely identified by its label l, while its motion state x = [px, px, py, py]T comprises of object s position and velocity in Cartesian coordinates. Each object moves in accordance with the constant velocity (CV) model given by xk = F CV xk 1 +q CV k 1. Here, F CV = [1, T0; 0, T0] I2, T0 is the sampling interval (T0 = 1 s for our experiments), denotes for the Kronecker tensor product; I2 is the 2 2 identity matrix; q CV k 1 N(0, QCV ) is a 4 1 zero mean Gaussian process noise, with co-variance QCV = σ2 CV [T 3 0 /3, T 2 0 /2; T 2 0 /2, T0] I2. The detection probability is p D(us, xp) = 0.98 ||xp us|| rd max(0, 0.98 (||xp us|| rd)ℏ) otherwise; where rd is the sensor detection range and ℏ= 0.008 m 1. The sensor reports false detections or false-alarm measurements following a Poison RFS with a clutter rate λ = 0.2, where each agent collects at most one measurement per time step for each object, either from the real objects, clutters (false detections) or the measurement is empty (missed detections). For sensor noise, the range and bearing based measurement is corrupted with a zero mean Gaussian process noise that depends on the distance between objects and agents, i.e., v N(0, R) with R = diag(σ2 φ, σ2 ρ) where σφ = σ0,φ + βφ||xp us||, σρ = σ0,ρ + βρ||xp us||; σ0,φ = 2π/180 rad, βφ = 1.7 10 5 rad/m, σ0,ρ = 10 m, and βρ = 5 10 3. Similarly, for vision-based sensor, each detected object x leads to a measurement z of noisy x y positions, given by: z = px, py]T +v. Here, v N(0, R) with R = diag(σ2 x, σ2 y) where σx = σy = σ0,xy + βxy||xp us|| with σ0,xy = 10 m, and βxy = 1 10 2. The grid size is 100 100 across four scenarios. This corresponds to a grid cell of 10 m 10 m for scenario 1,2 and 3 and a grid cell of 20 m 20 m for scenario 4. The total time is 200 s. The agent does not have any prior knowledge about object s state, thus it uses the initial birth probability r B = 0.005, and a Gaussian density p B = N(x; m B, QB) with m B = [500, 0, 500, 0]T and QB = diag([500, 10, 500, 10]). |