Need help with pytorch-PPUU?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

150 Stars 44 Forks MIT License 749 Commits 5 Opened issues


Code for Prediction and Planning Under Uncertainty (PPUU)

Services available


Need anything else?

Contributors list

# 8,023
300 commits
# 323,090
21 commits
# 95,223
1 commit
# 17,228
1 commit
# 759,232
1 commit

Prediction and Policy-learning Under Uncertainty (PPUU)

Gitter chatroom, video summary, slides, poster, website.
Implementing Model-Predictive Policy Learning with Uncertainty Regularization for Driving in Dense Traffic in PyTorch.


The objective is to train an agent (pink brain drawing) who's going to plan its own trajectory in a densely (stochastic) traffic highway. To do so, it minimises a few costs over trajectories unrolled while interacting with a world model (blue world drawing). We need to start, then, by training the world model with observational data from the real world (Earth's photo), which needs to be downloaded from the Internet.

Getting the real data

To get started, you need to fetch the real world data. Go to this address, and download the TGZ file (330 MB) on your machine. Open a terminal, go to the location where you've downloaded the file, and type:

tar xf xy-trajectories.tgz

This will expand the NGSIM (Next Generation Simulation) data set compressed archive, consisting of all cars trajectories for the 4 available maps (now 1.6 GB). Its content is the following:

├── i80
│   ├── trajectories-0400-0415.txt
│   ├── trajectories-0500-0515.txt
│   ├── trajectories-0515-0530.txt
│   └── trajectory-data-dictionary.htm
├── lanker
│   ├── trajectories-0830am-0845am.txt
│   ├── trajectories-0845am-0900am.txt
│   └── trajectory-data-dictionary.htm
├── peach
│   ├── trajectories-0400pm-0415pm.txt
│   ├── trajectories-1245pm-0100pm.txt
│   └── trajectory-data-dictionary.htm
└── us101
    ├── trajectories-0750am-0805am.txt
    ├── trajectories-0805am-0820am.txt
    ├── trajectories-0820am-0835am.txt
    └── trajectory-data-dictionary.htm

4 directories, 14 files

Finally, move the

directory inside a folder named

Setting up the environment

In this section we will fetch the repo, install the dependencies, and view the data we just downloaded, so that we can see if everything runs fine. So, open up your terminal, and type:

git clone [email protected]:Atcold/pytorch-PPUU.git
# or with the https protocol
# git clone

Now move (or symlink) the

folder inside the repo:
cd pytorch-PPUU
mv  .
# or
# ln -s 

Now install the

environment (this expects you have
on your system, go here if this is not the case):
conda env create -f environment.yaml
# To activate this environment, use:
# > source activate PPUU
# To deactivate an active environment, use:
# > source deactivate

As prescribed, activate it by typing:

source activate PPUU  # or
conda activate PPUU

Finally, have a look at the four maps available in the NGSIM data set, namely: I-80, US-101, Lankershim, and Peachtree. There is a "bonus" map, called AI, where I've hard coded a policy for the vehicles, which are using a PID controller. Type the following command:

python -map 
# where  can be one of {i80,us101,peach,lanker,ai}
# add -h to see the full list of options available

The frame rate should be greater than 20 Hz. Often it will be larger than 60 Hz. To be noted, here the vehicles are performing the actions extracted from the trajectories, and not simply following the original spatial coordinates.

Dumping the "state, action, cost" triple

In order to train both the world and agent models, we need to create the observations, starting from the NGSIM trajectories and the simulator. This can be done with the following command:

for t in 0 1 2; do python -map i80 -time_slot $t; done
# to dump the triple for the i80 map, otherwise replace i80 with the map you want

Upon the script termination, we will find a folder named

within our
. The content of the latter is now the following:
├── state-action-cost
│   └── data_i80_v0
│       ├── trajectories-0400-0415
│       │   ├── car1.pkl
│       │   └── ...
│       ├── trajectories-0500-0515
│       │   └── ...
│       └── trajectories-0515-0530
│           └── ...
└── xy-trajectories
    └── ...

Additional info

Each pickled vehicle observation is stored as

. Its content is a
which includes the items and corresponding sizes (shapes):
images               (309, 3, 117, 24)
actions              (309, 2)
lane_cost            (309,)
pixel_proximity_cost (309,)
states               (309, 7, 4)
frames               (309,)

For example, this vehicle was alive for 309 frames (time steps). The

represent the occupancy grid, which is as large as 4 lanes width (24 pixels, here).
  • The R channel represents the lane markings.
  • The G channel encodes the position and shape of the neighbouring vehicles.
  • The B channel depits our own vehicle.


is a collection of 2D vectors, encoding the positive and negative acceleration in both x and y directions. The
are the task specific costs (see slides for details). The
encode position and velocity of the current vehicle and the most closest 6 ones: left/current/right lanes, front/back. Finally,
tells us the snapshot time stamp, so that we can go back to the simulator, and inspect strange situations present in the observations.

Finally (this will likely be automated soon, and made avaiable for every map), extract the car sizes for the I-80 map with:


Training the world model

As we have stated above, we need to start by learning how the real world evolve. To do so, we train a neural net, which tries to predict what happens next, given that we start in a given state, and a specific action is performed. More precisely, we are going to train an action conditional variational predictive net, which resembles much a variational autoencoder (VAE) that has three inputs (concatenated sequence of

) and its output is set to be the next item in the sequence (

In the code, the world model is shortened as

, which stands for forward dynamics model. So, let's train the forward dynamics model (
) on the observational dataset. This can be done by running:
python -model_dir 

Training the cost model

Along with the dynamics model, we have a separate model to predict the costs of state and action pairs, which can be trained by running:


Training the agent

agent training

uncertainty computation

Once the dynamics model is trained, it can be used to train the policy network, using MPUR, MPER, or IL. These corresponds to:

  • MPUR: Model-based Policy learning with Uncertainty Regularisation (shown in the figure above)
  • MPER: Model-based Policy learning with Expert Regularisation (model-based IL)
  • IL: Imitation Learning (copying the expert actions given the past observations)

This is done by running:

python train_{MPUR,MPER,IL}.py -model_dir  -mfile 

Evaluating the agent

To evaluate a trained policy, run the script
in one of the three following modes. Type
to see other options and details.
python -model_dir  -policy_model  -method policy-{MPUR,MPER,IL}

You can also specify

-method bprop
to perform "brute force" planning, which will be computationally expensive.

Parallel evaluation

Evaluation happens in parallel. By default, evaluator script uses min(10, #coresavailable) processes. It doesn't go above 10 because then it hits GPU memory limits. To change the number of processes, you can pass

argument to `

script. Also, for this to work, you need to request cpu cores using
--cpus-per-task=X` argument for slurm. The slurm limits cpu usage to 64 cores per user, and gpus to 18 per user, therefore 3 is a reasonable limit to enable us to use all the gpus without hitting the gpu limit when running multiple evaluations. The CPU limit can be extended, but you need to email the IT helpdesk.

Pre-trained models

Here you can download the predictive model and the policy we've trained on our servers (they are bundled together in the

field of this Python dictionary). The agent achieves 82.0% of success rate.
Here, instead, you can download only the predictive models (one for the state and one for the cost), and try to train the policy by your own.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.