Final year project · 2025 to present
CASCADE
Collaborative Anticipatory Stacking with Collision-Aware Dynamic Execution. CASCADE explores proactive human-robot collaboration in assembly and stacking tasks. The system aims to understand human actions, anticipate upcoming intentions, coordinate task allocation, monitor assembly sequences, and adapt robot motion to maintain safe collaboration.
The problem
Collaboration that is more than reaction.
In most human-robot collaboration, the robot waits for a person to finish a step and then responds. Shared work becomes slow and stop-start, and the person and the robot may reach for the same part.
People avoid this by anticipating each other. CASCADE studies whether a robot can do the same: watch the human hand, estimate what comes next, choose a different component to work on, and keep the person safe throughout.
System architecture
Five stages from observation to safe motion.
The platform is a Franka Emika Panda arm with an RGB-D camera, controlled through ROS 2 and Python.
-
01
Perception
RGB-D camera and hand tracking observe the person and the components.
In progress -
02
Intention prediction
Sequence models (LSTM, Transformer) estimate the next human action.
Under investigation -
03
Task allocation
The robot picks a different component from the one the person will likely take.
In progress -
04
Sequence monitoring
Check the assembly order, detect wrong placement and restore the sequence.
Planned -
05
Safe execution
Distance-based safety zones adapt the speed and path of the Panda.
In progress
Four capabilities
What the system is designed to do.
Human Intention Prediction
RGB-D observations are used to study human hand movements and predict upcoming actions: reaching, grasping, transporting, placing, retracting and idle.
LSTM and Transformer models are under investigation. No accuracy figures are reported yet.
Anticipatory Task Allocation
The robot anticipates which component the human is likely to collect and proactively selects another suitable task, reducing competition for the same object.
Assembly Sequence Monitoring and Error Recovery
The system should monitor the required assembly sequence and detect incorrect component placement. The intended recovery is to remove or reverse the placement and restore the required sequence.
Collision-Aware Dynamic Execution
Distance-based safety zones adapt robot behaviour, implemented in ROS 2.
Dataset plan
A custom manipulation dataset.
Fine-grained, labelled human actions during assembly are needed to train and test intention models. The dataset is being built from RGB-D hand-tracking recordings.
| Participants (target) | 60 to 100 |
|---|---|
| Trials per participant (target) | about 20 |
| Action classes | Reaching, grasping, transporting, placing, retracting, idle |
| Models | LSTM, Transformer (under investigation) |
These are planned targets, not completed numbers.
Evidence
What is not claimed yet.
CASCADE is ongoing. At this stage I do not report:
- a finished dataset or its size,
- model accuracy for action or intention prediction,
- autonomous error recovery,
- safety validation of the distance-based zones with people.
I will update this page and the status labels as results become available.
Foundations
Built on work from the Human-Centered Robotics Lab.
- A 1 kHz Cartesian impedance controller on the same Franka Panda, supporting compliant physical interaction.
- A manipulation pipeline with Pinocchio, MoveIt and AprilTag perception, from simulation to hardware.
- A safety-filtered motion planner (GPMP2, MPPI and control barrier functions), verified in simulation.
Discuss this work
Questions or collaboration on CASCADE?
I am open to research opportunities and Master's supervision discussions in human-robot collaboration.