ZTE's EmbodiedMind: 8B model lifts long-horizon robot planning accuracy 30.31%
机器之心 · wechat · 2026-11-06
ZTE published the EmbodiedMind paper introducing a three-stage training pipeline (RSFT → IR-GRPO → Trie-GRPO). With only 522k training samples, an 8B model averages 70.02% across 18 embodied-brain benchmarks, improving long-horizon planning accuracy by 30.31% over prior best methods; the paper is submitted to ICRA 2027.
- RSFT: rejection-sampling stratifies trajectories, keeping only uncertain plus some easy samples to raise data information density
- IR-GRPO: per-task FIFO queues with moving-average quotas rebalance heterogeneous task sampling, fixing the seesaw effect
- Trie-GRPO: moves credit assignment from trajectory level to step level via an action-prefix tree, so correct intermediate actions in failed trajectories are no longer penalized
On real hardware, using an AgiBot A2 across 50 tabletop scenes, the model hits 62% on tidying-planning vs Qwen3-VL-8B's 40%, and scores 66.67 on a 25-step simulation benchmark with 60.45% success on ALFWorld. On SuperCLUE's September embodied-brain leaderboard, ZTE's 180B Nebula-EmbodiedBrain scores 77.48, tied for first domestically with Alibaba's Qwen3.8-Max, leading the embodied-safety sub-list at 97.92.
More from Embodied
- Galaxea Co-founder Hang Zhao's CoRL Keynote: Split Robot Intelligence into System 2/1/0 by Timescale — YuXiang_IRVL · 2026-11-13
- Apparent OpenAI hardware unboxing photo leaks: robot arm plus handheld console — yungcontent · 2026-11-13
- FINGR dexterous hand solves a 2x2 Rubik's Cube with continuous finger tricks, wins Best Paper at CoRL 2026 workshop — Scobleizer · 2026-11-13
- Fei-Fei Li amplifies new Atlas demo: 'It is very hard to explain how cool Atlas is' — drfeifei · 2026-11-13
- GXO to deploy up to 1,000 Apptronik Apollo humanoid robots in warehouses over four years — CyberRobooo · 2026-11-13
- Waymo robotaxi spotted quietly obeying a 'no right turn on red' sign humans often miss — reed · 2026-11-13