Description
In the second assignment, you will implement the predictive TTC forces approach for local navigation that we covered in lecture 2. The approach can be considered as a variant of the PowerLaw
model, as both approaches rely on forces that depend on the relative displacement of agents at the
moment of a collision. You will also incorporate uncertainty in the sensed velocities of the neighbors
using the isotropic formulation that we will cover next week.
In the zipped file ttc.zip, you will find three multi-agent navigation scenarios that you can use
to test your code. I have also provided Python code, that reads scenarios, and visualizes the agents.
Your task is to modify the Agent class in the agent.py file.
Basic Requirements
Your program should work as follows. It should first load a simulation scenario and set the time
step ∆t to be 0.05 s in the simulator.py file. Then, for each of the n agents you need to compute
a collection of forces that determine the behavior of the agent at each simulation step, and then
update its position and velocity. To do so
1. Please modify the computForces function in order to:
(a) Determine the goal force of the agent as explained in the class, i.e.
fg =
vg − v
ξ
(1)
Here, we assume that the goal velocity is the unit vector pointing from the current
position of the agent to its goal position scaled by the agent’s preferred speed (already
computed for you in the update function). A typical value for ξ is 0.5s.
(b) Determine all neighbors of the agent that are less than dH m away from the agent. A
typical value for dH is 5-10 units.
(c) For each neighbor estimate the corresponding time-to-collision value, τ , between the
neighbor and the agent. If there is a collision, add the corresponding repulsive force to
fg. You should refer to the lectures, but the force should be of the form:
fa =
max(τH − τ, 0)
τ
n (2)
where n denotes the unit vector that pushes the two agents apart at the moment of
impact. Typical value for the time horizon τH is 4s.
2. Modify the update function to update the velocities and positions of the agents based on
the computed forces by using simple Euler integration. You should cap the velocity of each
agent to its maximum speed. You may also want to cap the maximum amount of force that
is applied to each agent before computing the new velocity.
3. Once you have this code working, you should extend it to account for uncertainty in the sensed
velocities. In particular, you will implement the isotropic time-to-collision formulation that
we will cover next Monday. This approach works by assuming that the error between the true
and sensed relative velocities is smaller than some known constant . The uncertainty-aware
approach works exactly as the TTC forces model described above. The only difference is that
the time-to-collision routine needs to be slightly modified since some of the coefficients of the
quadratic include additional -dependent terms. In fact, if you set = 0, you should be getting
the vanilla TTT forces. Of course we need to have a model of the sensing error as well. In our
case, we will assume that at each time step the sensed relative velocity between two agents A
and B, ˆvAB , is the same as the true velocity, vAB .
Please test your code on the provided 3_agents, 8_agents, and crossing_agents scenarios.
Once you are happy with your results, you should export each simulation into a csv file by setting
the doExport flag to True in simulator.py. For each scenario, please export two simulation files,
one with = 0, and one with = 0.2. You should also document the parameters that you used for
each simulation in a Readme file. As opposed to project 1, a universal set of parameters should work
well here. Overall, you should be able to see much better results compared to the sampling-based
approach you implemented in the first assignment.
Extra credits (3pts)
Consider implementing the following extensions:
• Implement the PowerLaw model covered in lecture 6. The approach is very similar to TTC
forces. The only difference is the magnitude of the avoidance force given two interacting
agents. In TTC forces, the magnitude is inversely proportional to the time that it takes for
the two agents to collide, while in PowerLaw the magnitude is a bit more involved as it is
derived from the gradient of the PowerLaw energy function. See the lecture for more details.
Once more, for each scenario, please export two simulation files, one with = 0, and one with
= 0.2. You can set the exponential cutoff point, τ0 to either 0s or 3s.
• Modify the sensor error model for your isotropic uncertainty implementation by assuming
that at each time step the sensed relative velocity, ˆvAB , is randomly perturbed from the true
velocity, vAB , as follows:
ˆvAB = vAB + η,
where η is sampled randomly from a disk centered at 0 having radius ν. Please experiment
with different values of ν and , and submit results with = 0.2 and ν = 0.1. As long as
ν ≤ , you shouldn’t be seen any collisions.
• Implement the adversarial uncertainty model for the TTC forces and/or the PowerLaw. The
only difference to the vanilla approaches is to use ˆvAB −
xAB
kxAB k
instead of vAB upon computing
the repulsive forces and time to collision. Here, xAB denotes the relative displacement between
the two agents. The sensed relative velocity ˆvAB can either follow the noise model above or
be the same as vAB . Please submit results with = 0.2, and ν = 0.1 and/or ν = 0. For more
details see the lectures and here (Section III.B)
Submission
Submit the assignment using Canvas. You can work in pairs if you want to. If you do so, though,
please let us know in advance. Along with your code, please upload the three csv files corresponding
to the three scenarios that you simulated, as well as a Readme file indicating the parameters that you
ended up using to simulate each scenario. Since we are using multiple files, please make a directory
and add all files to it. You should zip your directory and upload it to the submission system.
Help
If you get stuck, please do not hesitate to contact us for help, and stop by during the office hours.
We also encourage you to post questions and initiate discussions on Canvas. Your colleagues are
also there to help you.