Summary
A Figure AI progress update on Helix (their humanoid VLA) applied to warehouse and logistics package handling. Reports per-package handling time down to ~4.05 seconds (~20% faster than a prior baseline), label-orientation-for-scanning accuracy up to ~95% from ~70%, and new robustness to deformable objects (poly bags, flat envelopes) rather than only rigid boxes — including learned auxiliary behaviors like patting down wrinkled mailers to improve barcode reads. Attributed to upgrades in “System 1,” the low-level visuomotor control policy, including “implicit stereo vision” for depth-aware motion. This is the vault’s first logged entry on Figure AI.
Key Contributions
- Quantified throughput/accuracy gains on a real deployed logistics task.
- Extension of humanoid dexterous manipulation to deformable (non-rigid) objects.
- A stereo-depth perception upgrade to the low-level control policy.
Strengths
- Real-world, large-scale deployed-task evidence (warehouse logistics) rather than a lab demo.
- Concrete before/after metrics for handling time and scan accuracy.
Weaknesses
- No architectural or training detail is disclosed publicly (proprietary).
- Numbers are self-reported by Figure with no independent benchmark.
- Exact publish date could not be pinned down with confidence.
Open Questions
- What specifically changed in “System 1” beyond “implicit stereo vision”?
- Is there any technical report accompanying this beyond the blog post?
Significance
Figure AI’s Helix line is one of the most closely watched humanoid VLA efforts in industry; this deployment update on deformable-object handling and throughput was a notable gap in this vault’s prior coverage.