The Trinity: capability-driven RSI lifts frozen-model benchmarks 34 points, robot success 0→60%
YangYou1991 · x · 2026-11-01
The Trinity introduces capability-driven multimodal self-improvement (RSI): agents build their own training environments and develop shared capabilities that transfer across tasks, all the way to real robots.
Key results:
- With model weights frozen: +34 points average across 28 training benchmarks, +26 across 18 unseen benchmarks
- Real2Sim2Real lifts Unitree G1 and AlohaMini2 real-robot task success from 0% to 60%
- Core mechanism: a sustained feedback loop between world generation and capability evolution
Project page, arXiv paper, GitHub code, and HF Papers are all publicly available.
More from Embodied
- SimReady 3D dataset with 1K+ assets trends on Hugging Face under CC-BY-4.0 — codepawl · 2026-11-01
- Shanghai startup Archon Robotics builds a Humanoid Foundation Model for whole-body intelligence — Scobleizer · 2026-11-01
- Robot Intelligence Index debuts: GPT-6 Astra tops robot model ranking at 1905 — haider1 · 2026-11-01
- RoboColiseum robot benchmark: Claude Opus 5.5 scores 76.28, edging out GPT-6 Astra — jiqizhixin · 2026-11-01
- ETH Zurich releases EgoHTR: 150k-frame egocentric 4D dataset for humanoid terrain traversal — Scobleizer · 2026-11-01
- Dyna's long-horizon robot demo hailed as one of the cleanest yet — Scobleizer · 2026-11-01