- LiDAR SLAM pipelines: feature extraction (edges/planes), deskewing & motion compensation, robust back-end (factor graph/iSAM2), loop closure (Scan Context/ISC), map management (submaps/voxels).
- Agentic tooling for robotics: reproducible eval harnesses, safe executors, dataset/bag orchestrators that don’t leak signals.
- Sensor fusion & drivers: LiDAR–IMU extrinsics, hardware-timestamp time sync (PTP/1588), ROS2 drivers (Livox/Ouster/Radar) under real-time QoS.
- Generative modeling: diffusion models, autoregressive LMs, controllable generation, efficient decoding strategies.
- Open LIO/VLIO stacks (CT-ICP/GICP/NDT, tightly-coupled preintegration, degeneracy handling).
- Place recognition & loop closure (Scan Context/M2DP/NDT-hist, outlier-robust relocalization).
- Calibration/time-sync toolchains (hand-eye, rolling-shutter modeling, multi-sensor clocks) + CI’ed bag benchmarks.
- Open-source LLM stacks (training & inference optimizations, fine-tuning methods like LoRA/QLoRA, MoE routing).
- Nonlinear optimization for SLAM: robust kernels, marginalization, observability, map priors.
- High-perf point-cloud ops: voxel hashing, surfel/TSDF mapping, CUDA-accelerated ICP/GICP.
- Driver-level engineering: zero-copy pipelines, rclcpp QoS tuning, clock discipline & timestamp hygiene.
- High-perf inference systems: CUDA/Triton kernels, tensor parallelism, KV-cache optimization.
- Reinforcement learning: policy gradient methods (A3C, PPO), value-based methods (DQN variants), model-based RL, and safe exploration strategies for real-world deployment.
- ICP/GICP/NDT/CT-ICP trade-offs and when each fails (degenerate geometries, dynamics).
- IMU preintegration in LIO, gravity alignment, extrinsics drift & how to bound it.
- ROS2 for LiDAR stacks: zero-copy, QoS profiles, bagging/replay that preserve timing.
- Transformer architectures: scaling laws, bottlenecks, tricks for training stability.
- Not just building models, but planting seeds of intelligence.
