AI

BridgeVLA++ Adds Spatio-Temporal Memory Without Blowing the Data Budget

Shar Hendrix 5 min read

Vision-language-action models are good at “do this now” and weak at “remember what you covered.” BridgeVLA++ (arXiv:2608.05042, submitted August 5, 2026) grafts a unified spatio-temporal memory onto the earlier BridgeVLA stack and claims you can keep the few-shot, 3D heatmap advantages while solving tasks that need history.

BridgeVLA++ teaser with benchmarks radar chart and real-world generalization panels
Project teaser: multi-view heatmaps, memory module, and benchmark scores. Source: bridgevla-plus.github.io.

What changed

BridgeVLA projects point clouds into multi-view 2D images and predicts intermediate heatmaps before robot actions, so the VLM backbone stays in its image comfort zone. BridgeVLA++ adds:

  • Temporal memory at the coarse stage (interaction history / keyframes)
  • Spatial memory at the fine stage (less-occluded initial geometry under the current zoom)

Both inject into patch-token space without rewriting the action heads. The project page says the memory adds about +9.2% parameters on a multi-billion-parameter backbone and roughly +0.22 s per step latency on an RTX 4090 in their timing setup.

Benchmark receipts (from paper + project site)

Headline success rates reported for BridgeVLA++:

SuiteScoreNote
RLBench (18 tasks)93.7%+6.9 pts vs prior SOTA cited
COLOSSEUM65.2%14 perturbation settings
GemBench51.1%four generalization levels
RMBench (dual-arm memory)96.0%base BridgeVLA 18.9% without memory
MemoryBench99.7%single-arm memory suite

Real-robot work spans a Franka Research 3 for general manipulation (BridgeVLA base reported 96.9% average with 10 demos per task; 95.4% with 3 demos) and a held-out Dobot CR5A for memory tasks, where BridgeVLA++ reaches 93.3% average on memory-dependent tasks in the basic setting versus 0–30% for the cited memory-free / weaker baselines on the same tables.

Architecture diagram for BridgeVLA and BridgeVLA++ memory injection
Architecture: 2D heatmap alignment plus temporal and spatial memory blocks. Source: project site.
Franka and Dobot real-robot experimental setups with ZED cameras
Real platforms used in evaluation (Franka FR3 and Dobot CR5A). Source: project site.

A Human’s Take

Memory is the difference between “pick the red block” and “uncover the block you covered two steps ago.” BridgeVLA++’s numbers on RMBench are the kind of jump that usually means the base model had no chance. I’m more interested in the Franka few-shot table than the absolute RLBench average — ten demos (or three) that still generalize under lighting and distractors is the practical bar for lab robots. Next check: open code and third-party replications on a different arm.

Sources