Context
I worked on multiple production-grade picking systems deployed in real industrial environments: general pick-and-place, total picking solution, order picking, and a Wonik deployment.
My Contribution
- Designed motion planners in C++ using OMPL
- Integrated perception, planning, and execution into a production pipeline
- Architected asynchronous task & motion sequencing
Technical details
The asynchronous sequencing framework decoupled perception, planning, and execution to improve throughput and hardware utilization.
System Stack
- Perception
- 6D object pose estimation
- Continuous scene updates for dynamic environments
- Task-Level Sequencing
- Finite state machine for pick → transfer → place → failure → replanning
- Asynchronous coordination between perception and planning
- Failure detection and recovery logic
- Planning
- Collision-aware motion planning
- Grasp candidate validation and selection
- Task-space trajectory generation for pre-grasp, grasp, and post-grasp motions
- Execution
- Low-latency command streaming to hardware
- Feedback-based correction during execution
- Safety monitoring with abort and fallback conditions
Impact
- Production deployment in factory environments
- Increased picking throughput and reduced failure cases
- Reduced per-cycle execution time by ~2 seconds
TODO: Add captions (task, hardware, throughput/latency improvement, and which deployment each image shows).