Simeon Nedelchev

A fast proprio-conditioned latent flow-matching planner that sits beside a frozen whole-body controller and turns high-level intent into trackable WBC references — reactive, physics-aware and controller-aware generation rather than open-loop playback. Lightweight enough for Gen and WBC to run together on the Unitree G1 onboard CPU (no Jetson), and generalizable across tasks under one unified WBC contract. Beyond locomotion teleoperation, the same planner can recover and stand up from failed states without a fall/stand mode FSM. On hardware: walk / sprint teleop, disturbance rejection, and agile recovery. Try Gen alongside WBC tracking to interactively generate robot locomotion (WASD to steer, Shift to sprint; press G to switch clips ↔ Gen). Training code and paper coming soon.

[Demo] [Deploy Code] [Simulation]

One shared MDP for whole-body motion tracking in mjlab — train once on a motion library, then deploy one policy across many skills. Supports multi-clip training, paper-specific task presets, and sim-to-real export.

[Project page] [Demo] [Code] [Docs]

Natural, robust humanoid locomotion with an offline predictor that maps lower-body state and velocity commands to upper-body and gait targets during RL training — while keeping deployment task-only on the Unitree G1.

Simeon Nedelchev, Ekaterina Chaikovskaia, Egor Davydenko, Eduard Zaliaev, Roman Gorbachev · 2026 preprint

[Project page] [Paper] [Code]

Senior engineer for RL locomotion and whole-body control on the Green humanoid at AIJ 2025: predictor-based human-like locomotion; a unified WBC for teleoperation, dance, and zero-shot skills; and motor and actuator system identification for MuJoCo simulation.

[Presentation]

Batched differential inverse kinematics on MuJoCo Warp with a Mink-shaped API: GPU-parallel task-space IK across many worlds — frame tasks, posture, COM, equality constraints for closed chains, and hard limits via GPU ADMM. Supports CUDA graph capture for real-time loops, with examples for Panda, UR5e, Cassie, dual iiwa, and G1.

[Docs] [Code] [PyPI]