Open-Source Harness Enables Qwen to Control Real Robots Cost-Effectively
Researchers from Tsinghua and UC Berkeley developed uni-robot-agent, an open-source harness allowing the Qwen3.8-Flash-Next model to operate real robots through tool calls without modifying model weights. This setup achieved performance competitive with closed-source systems at a fraction of the cost. This development demonstrates that general-purpose open-weight language models can effectively control physical robots without expensive specialized fine-tuning. By drastically reducing API costs compared to proprietary alternatives, it lowers the barrier to entry for robotics research and deployment. Using the uni-robot-agent harness, Qwen3.8-Flash-Next achieved an average API cost of $1.09 per task compared to $21.29 for GPT-6 Astra on real-world manipulation tasks. It also outperformed fine-tuned π₀.₅ models on specific real-world tasks without requiring robot-specific weight adjustments.
## BACKGROUND
Robot-Use Agents and Vision-Language-Action models represent cutting-edge approaches to connecting artificial intelligence with physical hardware, allowing AI systems to interpret visual data and execute physical manipulation tasks. Traditionally, controlling robots required expensive, highly specialized model fine-tuning or proprietary closed-source infrastructure.