Robotics: The Autonomous Manipulation Pipeline

The Autonomous Manipulation Pipeline

A systems-level exploration of how robots sense, plan, and act under uncertainty — spanning probabilistic localization, motion planning, and constraint-aware manipulation.

01. Probabilistic Self-Localization

Built an indoor localization system that enables a mobile robot to infer its pose from noisy motion and planar LiDAR data using probabilistic reasoning.

Indoor robots operate without GPS and must continuously reason about uncertainty. Small motion errors accumulate quickly, and raw sensor data is often ambiguous. This project explores how probabilistic inference can turn unreliable inputs into a stable estimate of where a robot actually is.

The Particle Filter Loop

Engineering Choices & Tradeoffs

• Implemented motion-dependent noise and a beam-based sensor model for robustness to dropouts and spurious readings
• Identified ~2,000 particles as an effective balance between accuracy and runtime performance
• Used K-Means clustering to extract a reliable pose estimate when the belief distribution became multi-modal

Seeing It in Action

Links

• Code: View GitHub Repository
• Technical Report (PDF): Read Report

02. High-Dimensional Motion Planning

Planned collision-free motion for a 6-DOF manipulator using RRT in configuration space, then improved trajectory quality with shortcut-based smoothing.

Robotic arms operate in high-dimensional configuration spaces, where even a simple “reach the goal” problem becomes a search over joint angles under strict collision constraints. This phase focuses on sampling-based planning for a 6-DOF manipulator and the practical steps needed to turn a valid path into usable motion.

Planning happens in joint space (angles), but targets are specified in task space (end-effector poses), so feasibility depends on consistent IK targets and collision-aware validation. I focused on making the planner not just find a path, but produce motion that would be reasonable to execute on a real arm.

The RRT Planning Loop

Engineering Choices & Tradeoffs

• Used goal biasing to accelerate convergence while keeping enough exploration to navigate around obstacles
• Implemented shortcutting / smoothing to reduce zig-zags and improve motion quality for actuators
• Tuned goal precision (ε) and introduced near-goal sampling to improve accuracy without exploding runtime
• Performed edge validation with incremental collision checking during extension to avoid adding invalid branches to the tree
• Compared raw versus smoothed trajectories to reduce waypoint count and joint chatter while preserving collision safety

Seeing It in Action

Links

• Code: View GitHub Repository
• Technical Report (PDF): Read Report
• Demo / Plots: (add link)

03. Constraint-Aware Manipulation

Executed precise, constraint-aware manipulation by defining valid task-space regions and controlling end-effector motion using Jacobian-based differential kinematics.

Grasping and manipulating objects requires more than reaching a point in space. The end-effector must approach objects from valid orientations and follow constrained paths to avoid slippage, collision, or instability. This phase explores how task-space constraints and differential control enable precise, physically meaningful manipulation.

Task-Space Control Loop

Engineering Choices & Tradeoffs

• Defined grasp constraints using TSRs to allow flexibility around cylindrical objects while enforcing orientation bounds
• Used the Jacobian pseudoinverse to convert Cartesian motion commands into coordinated joint updates
• Observed that large task-space steps violate the Jacobian’s linear approximation, introducing lateral drift
• Mitigated integration error by reducing step size and emphasizing incremental, closed-loop motion

Seeing It in Action

Physical execution of task-space constrained motion, demonstrating stable Jacobian-based control under real-world dynamics.

Links

• Code: View GitHub Repository
• Technical Report (PDF): Read Report

Contributors & Technical Scope