GR00T
GR00T is a technology serving as a backbone for Vision-Language-Action (VLA) robot policies. These models function by directly converting camera images and language instructions into robot motor commands.
In September 2026, a study analyzing physical failure modes of VLA models under camera faults evaluated the GR00T architecture alongside pi 0.5. The research found that while both image blackouts and freezing resulted in low task success rates, the physical behaviors differed significantly, with freezing causing extreme joint movements and blackouts leading to more frequent object drops.
In October 2026, the GR00T N.7 model was utilized in the HumanoidToolBench study, which introduced an 18-task benchmark for evaluating humanoid tool use. The research analyzed gaps between tool selection and task completion, noting that focused probes with GR00T N.7 showed reduced selection accuracy when encountering unseen tools and continued task execution even under unrelated instructions.
In September 2026, the DexRoam study used the GR00T N1.7 backbone to learn mobile bimanual manipulation from human demonstrations, improving average success rates from 29% to 56%. Additionally, the GeoAAC research published in September 2026 employed GR00T N1.5 as an evaluation target, confirming performance improvements over fixed-horizon baselines by dynamically adjusting the action horizon of VLA policies.
Key facts· every fact cites a source
| Camera fault failure mode study target | GR00T architecturesource: learning note — VLA Failure Modes Under Camera Blackouts and Freezes |
|---|---|
| HumanoidToolBench evaluation model | GR00T N.7source: learning note — HumanoidToolBench: Benchmarking Humanoid Tool Use |
| DexRoam backbone model | GR00T N1.7source: learning note — DexRoam: Learning Mobile Bimanual Manipulation from Human De… |
| GeoAAC evaluation model | GR00T N1.5source: learning note — GeoAAC Teaches Robot Policies When to Look Further Ahead |
Learning notes behind this entry4
- Oct 2 · paper · arXiv cs.ROHumanoidToolBench: Benchmarking Humanoid Tool Use
- Sep 30 · paper · arXiv cs.ROVLA Failure Modes Under Camera Blackouts and Freezes
- Sep 29 · paper · arXiv cs.RODexRoam: Learning Mobile Bimanual Manipulation from Human Demos
- Sep 18 · paper · arXiv cs.ROGeoAAC Teaches Robot Policies When to Look Further Ahead
Revision history4
·daily rewrite·local AI · qwen3.8 27b
Added the September 2026 study analyzing GR00T failure modes under camera faults.
from: VLA Failure Modes Under Camera Blackouts and Fre…, HumanoidToolBench: Benchmarking Humanoid Tool Us…, DexRoam: Learning Mobile Bimanual Manipulation f… +1 more
·daily rewrite·local AI · qwen3.8 27b
Added information about the use of GR00T N.7 in the HumanoidToolBench study for tool use evaluation in October 2026.
from: HumanoidToolBench: Benchmarking Humanoid Tool Us…, DexRoam: Learning Mobile Bimanual Manipulation f…, GeoAAC Teaches Robot Policies When to Look Furth…
·daily rewrite·local AI · qwen3.8 27b
Rewrote the summary from the remaining sources after 1 source note was withdrawn.
from: DexRoam: Learning Mobile Bimanual Manipulation f…, GeoAAC Teaches Robot Policies When to Look Furth…
·first version·local AI · qwen3.8 27b
First summary, written from 2 learning notes.
next scheduled rewrite (if new notes arrived):
A local AI model wrote this entry from the learning notes. If something is wrong or missing, tell us — the AI reads and answers your report first, and the operator reviews it; any fix shows up in the entry's revision history.