A research team led by Peng Ren and Kai Chen has introduced Persistent Object Tokenization (POT), a novel approach to closing the loop in humanoid Vision-Language-Action (VLA) policies by treating task objects as persistent physical entities across movement, contact, occlusion, and recovery.
The Object-State Divergence Problem
Long-horizon humanoid loco-manipulation presents a fundamental challenge: the object state used to condition a whole-body action can diverge from the state used to verify whether that action achieved the intended physical relation. Standard VLA policies treat objects as transient visual features, losing track of them during movement or occlusion.
How POT-VLA Works
The proposed POT framework maintains role-indexed 3D object records extracted from RGB-D observations. These persistent records are converted into object tokens that condition both the whole-body action expert and geometric predicate verification. This yields a closed-loop execution system where object state is simultaneously actionable and verifiable.
Real-World Results on Unitree G1
Instantiated as POT-VLA and tested on a Unitree G1 humanoid robot, the approach demonstrates dramatic improvements:
- 39/80 → 71/80 successes over eight real-world task families vs. a matched GR00T-N1.7 baseline
- 44/50 successes on Being-0-aligned service tasks, vs. 37/50 reported by the Being-0 paper
- Largest gains on tasks requiring maintained 3D spatial relations
Implications for Humanoid VLA
The results suggest that persistent, object-centered state representation is a useful abstraction for verifiable humanoid VLA execution. By grounding actions and verification in the same 3D object records, POT-VLA addresses a critical gap between action generation and outcome verification in long-horizon robot tasks.
Paper: arXiv:2607.18016 [cs.RO], submitted July 20, 2026.
