Geek+ shows Gravity embodied model: ~2s per pick, natural-language correction lifts success to 87.5%
机器之心 · wechat · 2026-11-03
At CeMAT 2026, Geek+ demonstrated its in-house Gravity embodied model powering the Gino1 humanoid and robotic arms: picking unfamiliar items at a stable 2 seconds per pick, with natural-language correction (e.g. "grab deeper") raising complex-task success from 17.5% to 87.5%.
Gravity highlights:
- Mixture-of-Transformers dual-brain architecture: cognitive side for task/environment understanding, action side for motion planning.
- Layered memory: short-term tracking of recent observations, long-term storing goals and execution stages for long-horizon tasks.
- Three efficiencies: async inference for tempo, zero-shot generalization to new SKUs without fine-tuning, autonomous recovery from anomalies.
Whole-warehouse play: Geek+ integrates embodied robots with its AMR fleet (RMS/WES, multi-robot scheduling) into end-to-end solutions, topped by a Warehouse Intelligence "warehouse brain." With 1,000+ customers and 81,000+ robots deployed, it also announced a 24/7 Autonomous Intelligent Warehouse plan targeting unattended, self-improving operations.
More from Embodied
- 'Beyond Validation Loss': best pretraining checkpoint may hurt downstream RL fine-tuning — SongShuran · 2026-11-04
- Liber-0 hits 36% zero-shot success on 60 robot tasks using 2.2M hours of human data — CyberRobooo · 2026-11-04
- DISCO distills tree search into policies, a CoRL 2026 Best Paper finalist for robot planning — yining_hong · 2026-11-04
- RoboColosseum Benchmark Finds Weak Cross-Embodiment Agreement in Robot Policies — huggingface · 2026-11-04
- DexSuite: open-source dexterous manipulation framework with 10 arms, 12 hands, 30 tasks — mangahomanga · 2026-11-04
- Fruit fly's 166,000-neuron circuit map is being used to drive robot behavior — CurieuxExplorer · 2026-11-04