Can Agentic Robotics Solve the Data Scaling Problem in Robotics?
AI agents use simulation to tackle tough robotics challenges with code, not data.
By Evan Ackerman, IEEE Spectrum Robotics Editor | 6 min read
!3D rendering of a SCARA robotic arm carefully dropping a syringe into a sharps box
A 3D rendering shows an agentic robotic arm depositing used hypodermic needles into a disposal box. In simulation, the task required agents to successfully navigate variable sub-challenges, such as differently shaped disposal boxes.
Source image: Kaiyuan Chen, Shuangyu Xie, et al. (Berkeley Automation)
Overview
Robotics has long struggled with a fundamental bottleneck: data. Unlike language models, which can learn from vast corpora of text scraped from the internet, embodied AI systems need physical interaction data—and that data is expensive, slow, and difficult to collect at scale. As of 2026, this "data scaling problem" remains one of the field's most persistent obstacles.
A promising new approach—agentic robotics—may offer a way around it. Rather than relying on massive datasets, AI agents operating in simulation are increasingly solving challenging robotics problems through code rather than raw data. By reasoning about tasks, generating and refining solutions programmatically, and iterating within simulated environments, these agents can produce capable robotic behaviors with far less reliance on costly real-world demonstrations.
Why Data Scaling Remains a Challenge
Traditional approaches to robot learning—imitation learning, reinforcement learning, and increasingly large-scale foundation models for robotics—all depend on training data. In the real world, collecting that data means teleoperation, motion capture, and countless hours of human demonstration. Even in 2026, with improved simulation pipelines and synthetic data generation, the gap between simulated and real-world performance (the "sim-to-real" gap) persists.
The core question driving current research: can intelligent agents substitute computation and reasoning for data volume?
The Agentic Approach
Agentic robotics reframes the problem. Instead of asking how to gather more training examples, it asks how autonomous AI agents can be deployed inside simulation to discover, test, and refine solutions to specific manipulation tasks. These agents can:
- Reason about task structure and decompose complex goals into manageable sub-tasks.
- Generate code that controls simulated robots, then iteratively debug and improve it.
- Generalize across variations—such as differently shaped objects or disposal containers—without requiring exhaustive labeled datasets.
In the example illustrated above, a SCARA robotic arm must deposit used hypodermic needles into a sharps disposal box. Rather than training on thousands of demonstrations, agentic systems can work through the task in simulation, adapting to variable sub-challenges like box geometry through programmatic reasoning.
Implications for 2026 and Beyond
As foundation models for robotics mature and simulation environments grow more realistic, the agentic paradigm is gaining traction. If agents can reliably solve tasks through code-driven reasoning in simulation—and transfer those solutions to physical hardware—the data bottleneck could become far less binding than it once was.
That said, significant challenges remain: closing the sim-to-real gap, ensuring safety and reliability in physical deployments, and validating that agent-generated solutions generalize beyond the simulated conditions in which they were developed.
Still, the trajectory is clear. In 2026, agentic robotics represents one of the most active frontiers in embodied AI—a potential path toward capable robots that learn not from endless data collection, but from intelligent, autonomous problem-solving.
Evan Ackerman is IEEE Spectrum's robotics editor.