EmbodiedLab Benchmarks Robots That Must Act to Know in Simulated Labs
HexaFuture · hf · 2026-10-27
A new benchmark from HexaFuture, Tsinghua, NUS and Peking University forces embodied agents to act to know: instead of stopping after state changes or substituting symbolic actions for experiments, robots must physically operate lab apparatus to discover scientific principles.
- 30 experiments across 5 disciplines on 14 apparatus platforms; system parameters are resampled per instance so outcomes can't be memorized but can be scored exactly.
- The agent reads instruments via cameras and chooses its own procedure; evaluation verifies claims are grounded in readings recorded during the episode.
- Six frontier models were evaluated through their developers' coding agents with native observation.
More from Embodied
- From 5-Word Prompt to Physical Product: AI Designs and 3D-Prints a Filter Holder — blixt · 2026-10-27
- NVIDIA opens ARIEL lab in Singapore, hiring agentic robotics research scientist interns — DJiafei · 2026-10-27
- He let Claude design a complete e-ink business card, from schematic to product page — isaul32 · 2026-10-27
- Zhejiang University's RoboHarness Lets LLM Agents Beat VLA Models on BEHAVIOR — zju · 2026-10-27
- VAST's EdenLive renders AI-generated 720p game worlds at 24 FPS on one RTX 5090 — jiqizhixin · 2026-10-27
- Tesla FSD safety data shows 9x longer gap between major crashes than manual driving — Scobleizer · 2026-10-27