Alec Bossard — Robotics & AI portfolio
Robotics · AI · Software Engineering. Toulouse, France. UPSSITECH SRI engineering student.
SVLR / RMBench
From a single observation to a persistent understanding of the world. Extending visual-language robotics for memory-dependent manipulation.
I connected mask-conditioned 3D perception, episode-scoped WorldMemory and modular execution. The language model selects symbolic actions; deterministic Python resolves geometry and motion.
End-to-end simulation integration and representative real-Franka demonstrations: Put Back Block, Rearrange Blocks, Press Button, Cover Blocks, and a tennis-ball placement. These physical demonstrations establish feasibility, not repeated hardware reliability.
Path Planning
An interactive 3D environment for autonomous racing trajectories, from a track centreline to an optimized racing line.
Path Planning Studio generates centrelines and optimized racing lines using RRT* shortcutting and quadratic-programming curvature minimization.
An interactive application for generating trajectories, comparing racing lines and inspecting simulation telemetry. Explore the original planner below.
UR3
Compute the motion of a six-axis robot, from a Cartesian path to joint trajectories.
Implemented UR3 forward and inverse kinematics, Jacobian-based differential models and Cartesian-to-joint trajectory generation.
Generates joint position, velocity and acceleration curves, then reconstructs tool position and velocity to inspect tracking errors. Dedicated model tests and optional PyBullet utilities accompany the pipeline.
GeoPolicy-Bench
Can a second camera help a learned robot policy keep placing a cube when its observations deteriorate?
Built a Panda simulation benchmark with behavior-cloning action chunks, calibrated point clouds, frozen evaluation protocols and replayable checkpoints.
V4: without augmentation, fusion placed successfully in 483/540 rollouts (89.4%), versus fixed-only 272/540 (50.4%), averaged equally over nine synthetic perturbation settings. The full test includes 2,400 rollouts across 12 policies and three training seeds. Tested augmentations showed no demonstrated overall gain.
Autonomous Robot
A mobile robot combining ultrasonic obstacle avoidance, wall following and color-based ball tracking.
Contributed robot motion and autonomous control, web HMI integration, and integration of the ball-tracking module with the control interface and robot behavior in a six-student team.
The repository includes recorded demonstrations of manual/autonomous driving, ball tracking, voice commands and perception/control integration.
EarlyEval
Predicting the outcome of an AI coding task from its first steps. An independent, reduced reproduction of EarlyEval.
Independently implemented a pipeline from observable execution prefixes to features, weighted prediction models, calibration and chronological replay. Evaluated two held-out agents, tested controlled ablations and documented what did not reproduce.
700 recorded runs across 350 shared tasks and two agents. Task 11 produced no early stops at the primary 0.95 threshold. At 0.75, it saved 42.62% of execution steps but distorted agent scores by 8.14 percentage points on average.