Tsinghua & Berkeley open-source uni-robot-agent: Qwen drives real robots at $1.09/task
Beneficial-Ad-8127 · reddit · 2026-11-04
Qwen developers from Tsinghua and UC Berkeley released uni-robot-agent, an open-source harness that connects general-purpose LLMs to real robotic systems via tool calls — no model weight changes or robot-specific fine-tuning.
Tested with Qwen3.8-Flash-Next running locally: across 19 real-world manipulation tasks, average API cost was $1.09 per task versus $21.29 for GPT-6 Astra; on the benchmark its score rose from 1.90 to 31.39, surpassing a previously reported GPT-6 Astra robotic-agent result; on real tasks it beat fine-tuned π0.5 by 5.75 points and zero-shot Dream-1 by 10. QwenDev officially amplified it, and the harness is open-source.
More from coding & agent
- Will code review disappear? Team already auto-approves AI-written PRs — Yuchenj_UW · 2026-11-04
- Using a Codex-built MCP skill to force YuE2 to generate instrumentals reliably — ActIcy7831 · 2026-11-04
- EnterpriseShift: a benchmark for role-based enterprise agent work with RBAC and async coordination — Shahules786 · 2026-11-04
- EnterpriseShift: a benchmark testing agents on role-based enterprise work across 13 jobs — Shahules786 · 2026-11-04
- NVIDIA intros NeMo Relay to inspect Hermes Agent traces — NVIDIAAI · 2026-11-04
- Agentic coding is commoditized—your judgment is all that's left — emax · 2026-11-04