Anirban Biswas

Designing and building a tensegrity-based humanoid robotic leg from scratch. Work spans CAD adaptation of joint mechanisms in SolidWorks/Fusion 360, 3D printing and CNC machining of components, procurement and integration of IMUs, encoders and force sensors, and adapting a ROS 2 motion controller for the new leg kinematics. Also handling Gazebo simulation and iterative physical testing.

Constrained Hardware Integration: Integrated IMUs, encoders, and force sensors into a live embedded control loop on a tensegrity humanoid leg system; managed real-world signal timing, noise, and hardware failure modes under operational constraints. Hardware Protocol Experience: Worked directly with CAN, I2C, and SPI at the hardware interface layer; diagnosed signal integrity issues and timing failures through structured hardware debugging. ROS 2 System Integration: Adapted and deployed a ROS 2 motion contro

On-Device Inference Pipeline: Deployed a real-time visual perception pipeline (camera input, Bi-LSTM classification, action output) on a resource-constrained embedded system with no cloud dependency, using ONNX Runtime optimised for low-latency on-device inference. Latency Optimisation: Profiled and optimised the full inference stack to meet real-time latency requirements under operational conditions; identified and resolved bottlenecks in preprocessing, model execution, and postprocessing stages. Model Hot

Full Build Cycle Ownership: Built, wired, and assembled dynamic physical robotic mechanisms from early-stage, evolving requirements through to live experimental validation. Avionics & Housing Packaging: Engineered sealed, ruggedized motor enclosures and spool mechanisms capable of sustaining extreme tension and impact loads; integrated sensitive encoders and IMUs into compact envelopes. Custom Wiring & Power Distribution: Built and routed custom wiring harnesses with strain relief; wired emergency power iso

Developing a Vision-Language-Action pipeline for adaptive human-robot interaction. Integrated a Vision-Language Model with OpenCV and MediaPipe-based perception and a transformer-backed Bi-LSTM classifier, running in ROS 2. Added Continual Learning (EWC + Experience Replay via Avalanche) so the system adapts to new users and environments without retraining from scratch. Optimised for real-time inference using ONNX Runtime.

Designed modular EV components in SolidWorks and Fusion 360 with a focus on design-for-manufacturing. Led hands-on wiring and structural assembly for 2- and 3-wheeler prototypes. Automated design workflows with Python, cutting prototyping cycles by 20%.