Research

Reinforcement learning methods for continuous control tasks have evolved in recent years generating a family of policy gradient methods that rely primarily on a Gaussian distribution for modeling a stochastic policy. However, the Gaussian distribution has an infinite support, whereas real world applications usually have a bounded action space. This dissonance causes an estimation bias that can be eliminated if the Beta distribution is used for the policy instead, as it presents a finite support. In this work, we investigate how this Beta policy performs when it is trained by the Proximal Policy Optimization (PPO) algorithm on two continuous control tasks from OpenAI gym. For both tasks, the Beta policy is superior to the Gaussian policy in terms of agent's final expected reward, also showing more stability and faster convergence of the training process. For the CarRacing environment with high-dimensional image input, the agent's success rate was improved by 63% over the Gaussian policy.

The CarRacing-v0 environment simulates an autonomous driving environment in 2D.

The observation space consists of top down images of 96x96 pixels and three (RGB) color channels. The latest four image frames were stacked and given as input to the agent's network after rescaling and preprocessing them to gray scale (totalling 84x84x4 input dimensions).
The action space has three dimensions: one encodes the steering angle and is bounded in the interval [-1, +1]. The other two dimensions encode throttle and brake, both bounded to [0, 1]. 

In the plot below, we can see the average rewards for 5 agents with different seeds for the CarRacing environment.
The final performance is shown on the bottom plots at a bigger scale, for agents with Gaussian policy (bottom left) and Beta policy (bottom right). 

Next, we see an illustration of the Gaussian and Beta stochastic policy distributions in relation to the action space of the CarRacing environment. 
For a fixed observation s (preprocessed image in left plot), we sampled the Gaussian and Beta policies for 5000 actions. 
For the Gaussian distribution, a significant portion of the actions fall out of the valid direction and brake/throttle range (both [-1, 1]), whereas for the Beta distributions, all actions fall within boundaries.

OpenAI CarRacing-v0 Leaderboard hosts a series of self-reported scores. We compare our results only to those found in peer-reviewed articles since they provide a basis for comparison and discussion.
Our proposal currently (as of 2021) surpass state-of-the-art models, beating all work reported previously in the literature.

Autonomous driving is a complex task, which has been tackled since the first self-driving car ALVINN in 1989, with a supervised learning approach, or behavioral cloning (BC). In BC, a neural network is trained with state-action pairs that constitute the training set made by an expert, i.e., a human driver. However, this type of imitation learning does not take into account the temporal dependencies that might exist between actions taken in different moments of a navigation trajectory. These type of tasks are better handled by reinforcement learning (RL) algorithms, which need to define a reward function. On the other hand, more recent approaches to imitation learning, such as Generative Adversarial Imitation Learning (GAIL), can train policies without explicitly requiring to define a reward function, allowing an agent to learn by trial and error directly on a training set of expert trajectories. % In this work, we propose two variations of GAIL for autonomous navigation of a vehicle in the realistic CARLA simulation environment for urban scenarios. Both of them use the same network architecture, which process high-dimensional image input from three frontal cameras, and other nine continuous inputs representing the velocity, the next point from the sparse trajectory and a high-level driving command. We show that both of them are capable of imitating the expert trajectory from start to end after training ends, but the GAIL loss function that is augmented with BC outperforms the former in terms of convergence time and training stability.

Images from the three frontal cameras located at the left, central, and right part of the vehicle, respectively. They were taken after the first few interactions of the agent in the CARLA simulation environment considering our defined trajectory. Each camera produces a RGB image with 144 pixels of height and 256 pixels of width. These images are fed to the networks as they are.

Architecture of the actor-critic network and discriminator - each of them has its own separate network, with the latter having an additional input for the action, in orange color, and a sigmoidal output $D(s,a)$ instead of the output layer of the actor-critic network which consists of the steering direction, throttle as actions for the actor (policy) and value of the current state $V(s)$ for the critic. The common, though not shared architecture (in blue) is composed of a convolutional block that process the images of the three frontal cameras, whose output features are concatenated with other nine continuous inputs for speed, next target point in the sparse GPS trajectory, and a high-level driving command. The resulting feature vector is input to a block of two fully-connected (FC) layers.

Average rewards vs environment interactions during training in the long route (setup 2). For each method (GAIL and GAIL with BC), the average performance of two runs (i.e, two agents trained from scratch) is shown with a stochastic policy (left plot) and a deterministic policy (right plot). The shaded area represents the standard deviation. The behavior cloning (BC) attains an average reward of 173.6 for ten episodes, while the maximum is at 760, achieved by both GAIL and GAIL augmented with BC.

Physics-informed neural networks (PINNs) impose known physical laws into the learning of deep neural networks, making sure they respect the physics of the process while decreasing the demand of labeled data. For systems represented by Ordinary Differential Equations (ODEs), the conventional PINN has a continuous time input variable and outputs the solution of the corresponding ODE. In their original form, PINNs do not allow control inputs neither can they simulate for long-range intervals without serious degradation in their predictions. In this context, this work presents a new framework called Physics-Informed Neural Nets for Control (PINC), which proposes a novel PINN-based architecture that is amenable to control problems and able to simulate for longer-range time horizons that are not fixed beforehand. The framework has new inputs to account for the initial state of the system and the control action. In PINC, the response over the complete time horizon is split such that each smaller interval constitutes a solution of the ODE conditioned on the fixed values of initial state and control action for that interval. The whole response is formed by feeding back the predictions of the terminal state as the initial state for the next interval. This proposal enables the optimal control of dynamic systems, integrating a priori knowledge from experts and data collected from plants into control applications. We showcase our proposal in the control of two nonlinear dynamic systems: the Van der Pol oscillator and the four-tank system.

MPC: Representation of the output prediction in a time instant, where the proposed actions generate a predicted behavior that reduces the distance between the value predicted by the model and a reference trajectory:

The PINC network has initial state y(0) of the dynamic system and control input u as inputs, in addition to continuous time scalar t. Both y(0) and u can be multidimensional. The output y(t) corresponds to the state of the dynamic system as a function of t 2 [0; T], and initial conditions given by y(0) and u. The deep network is fully connected even though not all connections are shown:

Below, modes of operation of the PINC network. (a) PINC net operates in self-loop mode, using its own output prediction as next initial state, after T seconds. This operation mode is used within one iteration of MPC, for trajectory generation until the prediction horizon of MPC completes (predicted output from the first Figure). (b) Block diagram for PINC connected to the plant. One pass through the diagram arrows corresponds to one MPC iteration applying a control input u for Ts timesteps for both plant and PINC network. Note that the initial state of the PINC net is set to the real output of the plant. In practice, in MPC, these two operation modes are executed in an alternated way (optimization in the prediction horizon, and application of control action).

a).            


b) 

Below, the representation of a trained PINC network evolving through time in self-loop mode (previous Figure a)) for trajectory generation in prediction horizon. The top dashed black curve corresponds to a predicted trajectory y of a hypothetical dynamic system in continuous time. The states y[k] are snapshots of the system in discrete time k positioned at the equidistant vertical lines. Between two vertical lines (during the inner continuous interval between steps k and k + 1), the PINC net learns the solution of an ODE with t \in [0; T], conditioned on a fixed control input u[k] (blue solid line) and initial state y(0) (green thick dashed line). Control action u[k] is changed at the vertical lines and kept fixed for T seconds, and the initial state y(0) in the interval between steps k and k + 1 is updated to the last state of the previous interval k  1 (indicated by the red curved arrow). The PINC net can directly predict the states at the vertical lines without the need to infer intermediate states t < T as numerical simulation does. Here, we assume that T = Ts and, thus, the number of discrete timesteps M is equal to the length of the prediction horizon in MPC.