VinRobotics Humanoid Robots Reaching Human-Level Walking Speed
VinRobotics engineers apply Reinforcement Learning (RL)–based locomotion to train humanoid robots for stable and natural movement, even on large and heavy platforms.
These advancements are built upon three core pillars:
1. Gait reward design (phase-aware, bio-inspired)
We redesigned the locomotion objective to include motion phase cues and human biomechanics priors, so the policy learns smooth, efficient natural steps, not just “go fast” or “don’t fall.”
2. Domain randomization for sim-to-real transfer
By owning the humanoid physical designs from scratch, we randomize key dynamic parameters. This makes the controller more resilient to modeling error and real-world variation, and improves real deployment reliability.
3. Curriculum learning
We train in a staged progression, gradually increasing difficulty while ensuring the policy masters each level of balance and control before moving on. This has been crucial for stability on a heavy platform.
This video features VR-H3, a high-end humanoid robot developed by VinRobotics, standing 178 cm tall and weighing 75 kg.
