Stride Forward: Pioneering the Future of Humanoid Robotics

Author(s): Luhui Hu

Originally published on Towards AI.

Humanoid robotics has always been at the cutting edge of artificial intelligence, merging intricate control systems with dynamic real-world challenges. In our recent work, STRIDE: Automating Reward Design, Deep Reinforcement Learning Training and Feedback Optimization in Humanoid Robotics Locomotion, we introduced a novel framework that not only automates the laborious process of reward design but also sets the stage for more agile, robust, and adaptive robotic systems.

Current Advances in Humanoid Robotics

Training a humanoid robot to walk, run, or balance is vastly different from teaching a simple robotic arm to move an object. Humanoid robots have dozens of joints, actuators, and sensors working in sync, creating an extremely high-dimensional control problem.

In deep reinforcement learning (DRL), the training process relies on reward signals, which shape the robot’s behavior over millions of simulated iterations. Designing an effective reward function is challenging because:

• Manual Reward Engineering is Slow — Defining rules for ideal movement is time-consuming and requires countless trials.
• Human Bias Limits Optimization — Manually designed rewards often favor human intuition rather than true optimality.
• Generalization is Difficult — A handcrafted reward function designed for one robot in one environment may fail in another.

Without efficient and scalable reward automation, humanoid robots remain constrained by static training methodologies.

STRIDE: A Paradigm Shift in Humanoid Robotics Training

Our framework, STRIDE (Structured Training and Reward Iterative Design Engine), automates the creation and optimization of reward functions, allowing humanoid robots to learn high-performance locomotion without human intervention.

How STRIDE Works

1. LLM-Powered Reward Generation — Using advanced large language models (LLMs) like GPT-4, STRIDE writes structured reward functions dynamically, eliminating the need for predefined templates.
2. Iterative Feedback Optimization — The framework continuously analyzes training outcomes and refines the reward function in a closed-loop manner.
3. Scalable DRL Training — With its optimized rewards, STRIDE trains robots to achieve sprint-level locomotion, surpassing traditional methods by over 250% in efficiency and task performance.

By removing manual reward engineering, STRIDE accelerates training cycles, enhances generalization across different robotic morphologies, and pushes humanoid locomotion to new heights.

Below is the STRIDE framework and please see the details in the paper.

Across diverse environments featuring humanoid robot morphologies, STRIDE outperforms the state-of-the-art reward design framework EUREKA, achieving an average improvement of round 250% in efficiency and task performance. Please see the comparisons between Stride and the SOTA Nvidia Eureka below:

You can find more detailed results in the paper if you are interested.

The Latest Advancements in Humanoid Robotics

Recent AI-powered humanoid robots have demonstrated stunning agility and dexterity, as seen in this recent YouTube showcase. Robots like Boston Dynamics’ Atlas and Tesla’s Optimus are proving that rapid advancements in AI, hardware, and control algorithms are making humanoid robots more viable in real-world settings.

Notable breakthroughs include:

• Parkour and Dynamic Motion — Atlas demonstrates advanced jumping, running, and climbing abilities using reinforcement learning and control optimizations.
• Dexterous Object Manipulation — Optimus showcases fine motor control, picking up and handling objects with increasing precision.
• AI-Driven Adaptability — Robots are beginning to self-correct and adjust to new environments without human reprogramming.

However, these systems still require heavily engineered reward functions — a limitation that STRIDE directly addresses.

How STRIDE Outperforms Existing AI Models

Most AI-driven humanoid robotics systems today rely on either:

• Manual reward design (slow and non-scalable), or
• Heuristic-based DRL training (lacking adaptability).

STRIDE outperforms existing models in three key ways:

1. Fully Automated Reward Generation

Unlike traditional methods requiring weeks of manual tuning, STRIDE leverages LLMs to generate high-quality reward functions instantly.

2. Continuous Self-Optimization

Whereas previous DRL methods rely on fixed rewards, STRIDE dynamically refines rewards based on training results, leading to faster and more stable learning.

3. Scalability Across Different Morphologies

STRIDE-trained reward functions generalize across different humanoid designs, making it a plug-and-play solution for robotics researchers and engineers.

The Future of AI-Powered Robotics

Looking ahead, STRIDE and similar frameworks will unlock next-generation humanoid robots capable of:

✅ Self-Learning and Adaptation — Robots that can learn new skills autonomously with minimal retraining.
✅ Advanced Human-Robot Collaboration — AI models that interact seamlessly with humans in daily tasks.
✅ Versatile Real-World Deployment — Robots transitioning from controlled lab settings to unstructured environments (factories, disaster zones, homes).

The Road Ahead

The STRIDE framework is not just an improvement in AI training — it is a transformational leap in the way we design, train, and deploy humanoid robots. By automating reward design, we eliminate a critical bottleneck, paving the way for AI-driven robots to move beyond rigid programming and towards true autonomy.

As humanoid robotics advances at an unprecedented pace, AI-powered optimization frameworks like STRIDE will be the key to unlocking their full potential.

🚀 Are you ready to stride forward into the future of humanoid robotics?

References

1. STRIDE: Automating Reward Design, Deep Reinforcement Learning Training and Feedback Optimization in Humanoid Robotics Locomotion: https://arxiv.org/abs/2502.04692
2. Eureka: Human-Level Reward Design via Coding Large Language Models: https://arxiv.org/abs/2310.12931
3. Top 10 NEW Humanoid Robots of 2025: https://www.youtube.com/watch?v=fBFwFyZbewo

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI