JAKUB
DWORAKOWSKI
HRI Researcher. Robot learning & interaction systems.
ABOUT
I explore the frontiers of human-robot interaction and adaptive robot learning in the wild. I build systems that perceive, act, and interact.
I am currently an HRI Researcher at Ludo Robotics.
Previously a Staff Roboticist, where I worked across the stack that lets a device understand people and act in physical space — from compact language-conditioned action models to active-perception policies trained with PPO, alongside state-estimation and motion-planning foundations.
Before that I spent three years in Special Projects on real-time perception: event-camera calibration, occupancy grids with hard latency guarantees, and lidar pipelines built on Apple Silicon. My roots are in mechatronics (Waterloo) and interactive intelligence (Georgia Tech).
Off the clock I prototype open interaction systems — most recently a voice-driven framework for the Reachy Mini, and a long thread of research on compressing and understanding neural networks.
EXPERIENCE
Ludo Robotics -- HRI Researcher
May 2026 — Present
Exploring the frontiers of human-robot interaction and adaptive robot learning in the wild.
Apple -- Staff Roboticist
Dec 2019 — May 2026
- Built a ~100MB language-conditioned action retrieval model for human-device interaction (~70% accuracy).
- Trained an active-perception policy in a continuous-action setting using PPO.
- Co-formulated a framework for human-device interaction with the design team.
- Produced a state-estimation stack for humans and objects using EKFs and discrete Bayes filters.
- Online controller tuning via genetic algorithm; SIMD 3×3 eigen solver; RRT expansion primitive with smoothness guarantees; binary-search collision checking (7× faster).
Apple -- Algorithms Engineer
Jun 2016 — Dec 2019
- Event-based camera calibration with real-time application.
- Hard real-time knot-a-not and convolutional cubic interpolation optimized for Apple Silicon.
- Occupancy-grid container with real-time memory-access guarantees and efficient CPU projection.
- Online lidar pipelines using Apple SIMD and Metal compute; header-only Metal/C++ compatibility library.
PROJECTS
Multi-Modal Interaction (Reachy Mini)
Voice-driven behavior planning (Whisper + GPT-OSS), JAX trajectory generation, open-set detection fused with EKF tracking.
Dynamic SVD Compression
Heuristic rank reduction during training that retains accuracy at roughly 20% of the original size.
Singular-Value Analysis
An SVD lens on gradient descent with/without momentum under dropout — plus a gradient-perturbation layer.
Depth-Enhanced ConvNet
Recovers RGB-D benefits without a depth channel; implemented in TensorFlow with transfer learning.
RL for PID Tuning
Custom continuous policy-gradient formulations for directly tuning controller gains online.
Neural-Driven Arm Control
Real-time closed-loop control of a live human arm with vanilla Q-learning, using a Kinect and electrode pads.
SAY HELLO
Let's build something that moves.
Open to research collaborations, consulting, and conversations about robot learning and interaction.