BridgeVLA++: A Data-Efficient, Generalizable, and Memory-Augmented Vision-Language-Action Framework for 3D Manipulation

TL;DR

BridgeVLA++ extends the prior BridgeVLA 3D VLA (ByteDance Seed + CASIA + NLPR) with a unified spatio-temporal memory architecture, addressing three limitations of existing 3D VLAs: data-hunger, brittle OOD generalization, and lack of history-of-observations. The base recipe — project raw point clouds into multi-view images, predict intermediate heatmaps aligned with a pretrained VLM’s input–output space, then decode robot actions — is preserved so the pretrained-VLM alignment (BridgeVLA’s data-efficiency lever) survives. Adds only ~9.2% parameters for the memory module and reports state-of-the-art across five manipulation benchmarks including two challenging memory-dependent ones, while claiming to preserve BridgeVLA’s data efficiency and OOD generalization. Also validated on bimanual settings and a new real-world platform.

Key claims

3D VLAs today are simultaneously data-hungry, OOD-brittle, and memory-less; these three deficiencies rule out data-scarce, open-world, and memory-dependent manipulation regimes and motivate an architecture that fixes all three without sacrificing any [§1].
BridgeVLA (v1)‘s core lever is input–output alignment preservation — raw 3D point clouds are projected into multi-view images and the model predicts intermediate 2D heatmaps before decoding actions, so the pretrained VLM sees inputs and produces outputs in its natively-trained modality [§1, §2].
BridgeVLA++ adds a unified spatio-temporal memory module on top of BridgeVLA v1 that models both persistent spatial context and temporal interaction history in one architecture — as opposed to separate spatial and temporal memory heads [§1, §3].
The memory module costs only ~9.2% additional parameters relative to BridgeVLA v1 [§1, abstract].
Reports state-of-the-art across five manipulation benchmarks, including two challenging memory-dependent manipulation benchmarks, without sacrificing the base model’s data efficiency and OOD generalization [§4, abstract].
Framework transfers to bimanual manipulation and to an additional real-world robotic platform — the paper claims scalability across tasks, environments, and robot platforms [§4, §5].

Method

BridgeVLA (v1) starts from a pretrained VLM and forces its inputs and outputs to stay in the VLM’s native modality: raw point clouds are projected into a set of multi-view RGB images, the model predicts intermediate heatmaps (2D on those views) as an aligned intermediate representation, and only then are 3D robot actions decoded from the heatmaps. This keeps the fine-tuning distribution close to the VLM’s pretraining distribution, which is what gives v1 its data efficiency and OOD generalization on 3D manipulation. BridgeVLA++ preserves this pipeline verbatim and inserts a unified spatio-temporal memory architecture between the multi-view perception and the heatmap prediction — the same module represents both persistent spatial context (what has been seen in the scene) and temporal interaction history (what the robot has done and observed over past steps). The memory module is a lightweight add-on (~9.2% params), so the pretrained VLM’s forward pass remains dominant; the paper positions the memory as augmenting the aligned pipeline rather than replacing any component of it.

Results

Headline claim from the abstract: state-of-the-art across five manipulation benchmarks, including two challenging memory-dependent manipulation benchmarks, at only ~9.2% additional parameters over BridgeVLA v1 [§4, abstract]. Data efficiency and OOD generalization of the base model are reported as fully preserved or improved rather than traded off for memory capability [abstract]. Bimanual and additional real-world platform results are reported to establish cross-embodiment scalability, but the abstract does not surface per-benchmark numbers; specific benchmark identities and success rates are deferred to the paper body (not extractable at filing).

Why it’s interesting

Directly answers the “how do we add memory to a VLA without breaking the base recipe” question that RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark posed with its π0.5 21.5%-TSR baseline — but stakes out an architectural memory position (unified spatio-temporal module inside a 3D-aligned VLA at ~9.2% param cost) that contrasts with PrediMem’s external-keyframe-bank + predictive-coding-auxiliary-head recipe on the same benchmark family. Sits alongside the “memory-into-weights via TTT” position of RoboTTT — 8k-timestep robot policy via test-time training (Jim Fan / NVIDIA GEAR) and the “in-context system identification” position of In-Context World Modeling for Robotic Control — three structurally different answers now on the board for VLA long-horizon memory. Also sharpens the VLA Models recipe-lever board by adding a 3D-aligned input pipeline + memory-augmentation recipe distinct from the existing pointing/action-pretraining/clean-teleop/frozen-WFM/UMI-scale/native-video-pretraining levers — the “keep the VLM’s input–output distribution intact” bet is closest in spirit to See like a Robot: Robot-Centric Pointmaps for Vision-Language-Action Models (robot-frame 3D pointmaps as input) but preserves 2D heatmaps as the aligned intermediate rather than adding a new input encoder.