Motivation
Brain–computer interfaces (BCIs) have the potential to restore independence for individuals with paralysis by translating brain activity into commands for computers and robotic devices. However, these decoded movement commands can be noisy or imprecise, making everyday tasks such as reaching, grasping, and eating difficult.
Shared autonomy offers a potential solution by combining user-generated movement commands with computer assistance. Rather than relying entirely on the user's input, the system estimates what the user is trying to accomplish and provides assistance toward that target.
Our Solution
Our solution centers on BLENT, a shared-autonomy control algorithm designed to infer a user's intended target from their movement commands and assist them in reaching it.
BLENT uses Bayesian intent inference, a method that updates the probability of each possible target as the user moves. It compares the user's movement direction with the direction of each target, assigning higher probabilities to targets that better match the observed movement. These probabilities are then used to calculate an assistance direction, which is combined with the user's original command to guide the cursor. BLENT does not receive the user's true intended target. It relies entirely on their movement commands, the cursor position, and the locations of possible targets. The amount of computer assistance can be fixed or adjusted according to the algorithm's confidence in its prediction.
(See the video at the bottom of this page!)
The broader goal is to integrate this approach with brain-controlled robotic devices to help individuals with tetraplegia regain independence in everyday activities, including eating and drinking.
My Contributions
I designed and implemented BLENT in Python, developing its target-inference and movement-assistance methods to combine user control with computer-generated guidance. I also built a modular testing framework supporting direct human control, simulated movement, artificially added noise, and shared autonomy. I used a SpaceMouse, a handheld controller that detects movement in multiple directions, as a substitute for neural control signals to evaluate how BLENT handled imprecise input in 2D cursor and 3D virtual-reality tasks.
I evaluated the BLENT algorithm across 132 combinations of assistance and noise levels (11 assistance levels × 12 noise levels), conducting 100 trials per condition for a total of 13,200 trials. I developed automated tools to record and analyze movement accuracy, path efficiency (how directly the cursor traveled toward its target), and task completion time, identifying a 70% fixed computer assistance as the best-performing setting among those tested.
Project Outcomes
Poster(s):
InterfaceNeuro 2026 Poster
Looking to discuss further? Contact me at research@mkmaharana.com
