From an ambitious pet feeder to a robust autonomous retrieval system. This project was a lesson in hardware reality, requiring us to pivot, adapt, and engineer creative solutions to bridge the gap between theory and the physical world.
While we ultimately pivoted away from the granular "Pet Feeder" application due to gripper limitations, our final "Fetch" implementation successfully met all the core technical design criteria. The robot demonstrated autonomous perception of a target, smooth trajectory planning using Bézier curves, and multi-stage manipulation to interact with the environment.
The system proved that the Innate Mars platform, despite its documentation gaps and hardware quirks, is capable of complex tasks when governed by a robust software architecture.
Given the timeline constraints, several "hacks" were implemented to ensure stability. With additional time, we would address these technical debt items to create a truly production-ready system.
Current Flaw: Our vision system relies on
simple color thresholding and plane filtering. This makes it
sensitive to lighting changes.
Improvement: Implement machine learning-based
object detection (YOLO) and integrate the LiDAR sensor to fuse
depth data, allowing detection of non-white objects in complex
backgrounds.
Current Flaw: The "Return to Base" feature
currently relies on simple odometry tracking (reversing distance
d), which accumulates drift over time.
Improvement: Implement SLAM (Simultaneous
Localization and Mapping) to allow the robot to navigate back to
a specific (x,y) coordinate map, regardless of the path taken.
Current Flaw: The arm moves to a pre-calculated
coordinate and grabs blindly. If the robot slips during the
approach, the grasp fails.
Improvement: Implement "Visual Servoing"
(closed-loop control), where the arm camera continuously
corrects the hand position during the descent phase.