Simeon Nedelchev

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

[Demo] [Code] [Docs]

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

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

[Project page] [Paper] [Code]

Combines learned motion matching and flow matching for online, command-conditioned full-body locomotion. Built for Unitree G1 but portable across robots: MuJoCo-grounded state, velocity commands (extensible to waypoints and trajectories), and interactive Viser control with seamless stand / walk / run transitions. Inference at ~800 FPS (LMM) and ~150 FPS (flow); PoC training from a handful of LAFAN clips. Walking and tracking validated with BeyondMimic / ZEST in simulation and on hardware.

Ongoing · code not yet published

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

[Presentation]

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

[Docs] [Code] [PyPI]