New VLA papers target finer robot manipulation
W2-VLA and BridgeVLA++ add wrist forecasting and memory for 3D robot manipulation.
Why it matters
Both works address gaps in current VLA systems for robotics: handling fine-grained local interactions, distribution shifts, data scarcity, and memory-dependent tasks. The papers point toward manipulation models that use task context and temporal history more explicitly rather than treating all visual inputs as equivalent snapshots.
The key points
- 1.W2-VLA forecasts future wrist latents for action prediction.
- 2.W2-CoT supplies structured manipulation annotations for supervision.
- 3.BridgeVLA++ adds spatio-temporal memory to BridgeVLA.
Two new papers describe vision-language-action frameworks aimed at improving robot manipulation. W2-VLA models future wrist-view latents conditioned on task context and wrist history, using a W2-CoT annotation pipeline for auxiliary supervision. BridgeVLA++ extends BridgeVLA with unified spatio-temporal memory to reason over observation histories while preserving the earlier framework’s data-efficiency and generalization approach.
⚡ Try this today
Read both papers before designing VLA manipulation systems that rely on wrist cameras, point clouds, or observation history.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
Enjoyed this brief? Get the next one in your inbox.
More in Research
Prior Audit-Repair Context Shifts LLM Verifier Thresholds Toward Leniency
arXiv:2608.
Improving the matrix multiplication exponent with modern optimization and AlphaEvolve
arXiv:2608.
[Paper] Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
We introduce Mobius-v0, an architecture that comprises a globally shared Memory (FFN) that stores knowledge vectors and multiple Reasoners (Self-Attn) that iter