S-Agent 2.0: a toolkit that gives coding agents 3D spatial reasoning, no retraining needed
liuziwei7 · x · 2026-10-25
Researchers from NTU and Ropedia released S-Agent 2.0, a spatial reasoning toolkit that turns a general-purpose coding agent into a system that reconstructs, measures and reasons about 3D environments — no model retraining required. It fixes failures where agents misread reference frames or ignore measurements in favor of visual intuition, by adding perception commands with explicit reference frames and scene memory to the agent's existing loop. Scores: 74.1 on MMSI-Bench, 90.5 on ReVSI with a Gemini-3.8-Flash backbone; +20.5 toolkit gain on ReVSI for Qwen3.8-27B (61.8→82.3), ahead of SpatialClaw on 5 benchmarks.
More from coding & agent
- Open-Source Secret Management for Agents: Egress Proxy in Production — ccerrato147 · 2026-10-25
- Which agent-stack tool would you replace first? A practitioner's keep-or-kill review — Cold_Hall_5384 · 2026-10-25
- Clip Control: a $79 one-time MCP server that clips long recordings with auditable receipts — KindDocument3016 · 2026-10-25
- Open-source fix stops Google Antigravity IDE from resetting its window to 1400x900 — Greedy-Permit-2999 · 2026-10-25
- CASPA summit: chip-to-chip interconnect eats 60-70% of package energy, AI chip design targets 9-month cycles — ai · 2026-10-25
- Engineer orchestrates agents from his phone by voice: 'engineering has never been easier' — KevinNaughtonJr · 2026-10-25