Robots can now imitate human actions with impressive accuracy, but they struggle to fix their own mistakes in unpredictable real-world environments. That's the gap that Dr. Sun Peng, a reinforcement learning expert who previously led AI research at ByteDance and Tencent, is now working to close at Stardust Intelligence.

The challenge facing the robotics industry reveals a fundamental truth about how machines learn. Imitation learning, the dominant approach today, teaches robots "how to do it" by copying human demonstrations. But once a robot encounters a situation it hasn't seen before, or makes a mistake, it often has no way to recover. That's where reinforcement learning comes in: it teaches robots to learn from their failures and continuously improve through environmental feedback.

Why Does This Matter for Robots Moving Into the Real World?

The robotics industry is at an inflection point. In laboratory settings, robots are evaluated on whether they "can complete a task." But when robots move into factories, warehouses, and homes, the evaluation criteria shift dramatically. The real question becomes: "Can it do the task stably and repeatedly?".

This distinction is crucial. A robot that succeeds 60% of the time in a controlled environment might fail catastrophically in a real factory where unexpected obstacles, equipment variations, and environmental noise are constant. The ability to recover from errors, adapt to long-tail interference, and make autonomous decisions over extended periods determines whether a robot has genuine commercial value.

Dr. Sun Peng brings a decade of experience translating this principle across different domains. At Tencent, he developed deep reinforcement learning systems that trained wheeled robots to follow moving targets autonomously. At ByteDance, he led the development of ByteRL, a large-scale reinforcement learning infrastructure that powered multiple game-playing AI agents, including a Hearthstone AI competitive enough to beat top professional players.

How Does Stardust Intelligence Plan to Integrate Reinforcement Learning Into Robot Development?

Stardust Intelligence has built what it calls a "full-stack technical system" designed around the principle that a robot's body, data, and AI models should work together from the start, not in isolation. The system consists of three interconnected layers:

• Base Models (Lumo Series): These AI models handle the robot's understanding, prediction, and generation of environmental actions. Lumo-1 teaches robots to understand the intent behind actions, not just the mechanics. Lumo-2 introduces a latent world dynamics approach, allowing robots to "think first, then act" by reasoning about future states before taking action.
• Front-End Agent (Philia): This layer manages user-facing capabilities including long-term memory, task management, multi-robot collaboration, and natural language interaction.
• Reinforcement Learning Post-Training: This is where Dr. Sun Peng's expertise becomes critical. It's the bridge connecting what the AI model can theoretically do with what the robot can actually do reliably in the real world.

The reinforcement learning component solves a specific problem: how robots can continuously learn and optimize their behavior through repeated interaction and task feedback in real environments. This is fundamentally different from training a robot once and deploying it. Instead, the robot improves with every execution.

"Over the past ten years, I have been working on reinforcement learning and agents, and have personally experienced the process of reinforcement learning moving from robot control and complex games to post-training of large models. This logic has been verified in the digital world, and now I want to bring these accumulations back to the physical world," said Dr. Sun Peng.

Dr. Sun Peng, Former Head of Agent Center at Tencent Robotics X

Dr. Sun Peng's statement reflects a broader shift in AI development. The techniques that made large language models more capable and aligned with human intent, particularly reinforcement learning from human feedback (RLHF), are now being adapted for physical systems. The core insight is the same: agents learn better when they receive feedback and can adjust their strategies accordingly.

What's the Practical Bottleneck Holding Back Robot Deployment?

Industry consensus is crystallizing around a key insight: after a robot reaches about 60% task success through imitation learning alone, the marginal benefit of further optimizing action reproduction drops sharply. The real bottleneck becomes the robot's ability to adapt to open, unpredictable environments and recover from exceptions.

This explains why so many humanoid and task-specific robots remain in pilot phases despite impressive demonstrations. They work well in controlled conditions but struggle with the variability of real deployment. Reinforcement learning addresses this by enabling continuous self-improvement based on real-world performance data.

Dr. Sun Peng's background demonstrates the scalability of this approach. At ByteDance, he expanded his work from individual robot control to large-scale distributed reinforcement learning systems capable of training multiple AI agents simultaneously. His team's Hearthstone AI achieved competitive-level performance against professional players, a benchmark that required the system to handle complex, dynamic decision-making.

The hiring of Dr. Sun Peng signals that Stardust Intelligence is preparing for the next phase of embodied AI development. Rather than focusing on whether robots can perform tasks in isolation, the company is building infrastructure for continuous learning and improvement in deployed systems. This represents a maturation of the robotics industry from prototype-focused development to production-ready systems that improve over time.

As robots transition from laboratories to real-world environments, the ability to learn from mistakes and adapt to unexpected situations will likely become the defining competitive advantage. Dr. Sun Peng's expertise in building large-scale reinforcement learning systems positions Stardust Intelligence to address this critical gap in embodied AI development.