3DSceneBench tests whether coding agents can build functional 3D worlds
yizhu59 · x · 2026-11-14
Prana Labs and UC Berkeley launched 3DSceneBench, a benchmark asking whether coding agents can build 3D worlds that actually work — not just look right.
- Scale: 49 whole-room scenes across 7 categories, up to 3,002 objects per scene, 471 tasks and 36,871 per-object tests.
- Setup: the agent gets 8 posed photos, up to 16 close-ups, and plain-language tasks (no weights/friction given), then writes one MuJoCo scene within 180 minutes or 10M output tokens — every object a named body, every door and drawer a real joint.
- Scoring: verifiers load the world, render and measure it, push objects, open doors, and run a robot inside, scoring appearance, geometry and functionality against the reference room. In one example, a submitted world's car rolled only 4.7cm under the same push.
- Paper, code, data and a leaderboard are all public.
Related event: 3DSceneBench Tests Whether Coding Agents Can Build Functional 3D Worlds(2 posts)→
More from coding & agent
- Val Town kills *.val.run links on Jan 10, 2027; adds MCP tools to set up redirects — threepointone · 2026-11-14
- Transcript Testing: Apache DataFusion maintainer says AI-era tests should be written for humans to read — mgill25 · 2026-11-14
- Scope launches Competitors to track how AI agents evaluate and bypass your product — RaphaelDabadie · 2026-11-14
- Open-source IMAP MCP lets Claude read and send email from any mailbox, with OAuth and hard safety rails — Ordinary_Ad_404 · 2026-11-14
- Claude Code 2.1.325 is about to be released — ClaudeCodeLog · 2026-11-14
- Grok Bot joins Google Meet, Zoom and Teams as a working meeting teammate — XFreeze · 2026-11-14