AI - Trajectory Modeling

🛫Predicting Trajectories with Directed-Info GAIL (2020-2021)

Abstract:

A VAE‑pretrained Directed‑Info GAIL Imitation Learning algorithm (based on Reinforcement Learning) for predicting and generating aircraft and robot trajectories from initial states to final goals, using enriched latent and environmental features to produce complete sequence policies.

Thesis Abstract:
As noted in the Directed-Info GAIL paper “the use of imitation learning to learn a single policy for a complex task that has multiple modes or hierarchical structure can be challenging”. This thesis will explore the use of Directed-Info GAIL algorithm, which is based on the generative adversarial imitation learning framework to automatically learn subtask policies from unsegmented demonstrations of robot trajectories and aircraft trajectories, given that flights and robots have indeed different modes of behaviour in different segments of trajectories, depending on tasks they fulfil and many trajectories’ contextual features.
Documentation DOI/Link: http://dx.doi.org/10.26267/unipi_dione/1112

Description:

A research system applying Directed‑Info GAIL to learn trajectory modes and policies for robots and aircraft from unsegmented demonstrations, generating complete trajectories from initial states to final goals.

✈️ Aviation Domain:
The model incorporates meteorological features in 3‑latent and 5‑latent variable experiments using real historical flights between Barcelona and Madrid. A Variational Autoencoder (VAE) is used for pre-training to capture latent trajectory modes before feeding into GAIL, which generates full trajectories (~1,000 timestamps) from initial states (longitude, latitude, altitude) to final goal states. Trajectories are visualized in QGIS maps, with different modes color-coded for clarity.

🤖 Robotics Domain:
The system models trajectories for the Hopper robot in OpenAI Gym / MuJoCo, a single-legged hopping robot. Input features include yaw, joint positions and velocities, actuator signals, torso orientation, linear and angular velocities, and contact sensors. A VAE captures latent movement modes before GAIL optimizes the policy. The simulation uses Python for environment interaction, MuJoCo (C/C++) for physics, and OpenGL for 3D rendering. Full trajectories are generated from initial states to motion goals, capturing realistic hopping behaviors over time.

🌊OceanVoyagerAI-Predicting Trajectories with VAE and Directed-Info GAIL 🚢 (2025)

Abstract:

An imitation learning algorithm predicting cargo vessel trajectories using VAE and Directed-Info GAIL on a custom maritime dataset.

Description:

An imitation learning algorithm for cargo vessel navigation, combining Variational Autoencoders (VAE) with Directed-Info GAIL, trained on a custom-generated maritime dataset. The system models custom realistic-like vessel trajectories, learning subtask policies for complex navigation scenarios and enabling trajectory prediction under multiple environmental and contextual conditions.