BridgeVLA++ Bolts Memory Onto 3D VLA, Hits 93.7% RLBench

The interesting thing in the new BridgeVLA++ paper on Hugging Face is not the leaderboard number, it is the shape of the fix. The authors, from the Chinese Academy of Sciences and ByteDance Seed, took their earlier BridgeVLA model, kept the frozen PaliGemma backbone, and bolted on a spatio-temporal memory module of 269.77M parameters, a 9.2% overhead on a 3B backbone. That is a small addition doing a lot of work.

The headline results are that BridgeVLA++ reports 93.7% average success on RLBench, up from 90.5% for the base model, and 65.2% on COLOSSEUM under distribution shifts. The more telling number is RMBench, the memory-dependent dual-arm benchmark, where the memoryless BridgeVLA scores 18.9% and BridgeVLA++ scores 96.0%. On the single-arm MemoryBench it reports 99.7%. The pattern is that adding a temporal memory of anchor views, neighboring keyframes, and adaptively selected sub-goal keyframes, plus a spatial memory that re-renders the initial point cloud under the current fine-stage camera, closes the gap on tasks that require actually remembering what happened earlier in the episode.

Why this matters if you are not training VLAs yourself: a lot of robotics work over the last two years has been about scaling the base model. This paper argues that for memory-dependent manipulation you can leave the backbone alone and get most of the way there with a modest, targeted addition. That has implications for cost and for the number of teams that can plausibly play in this space.

The caveats are worth stating plainly. Inference goes from 0.35 seconds per step to 0.57 seconds per step on an RTX 4090, which is not free for reactive control. The slot budget for cached keyframes is tuned per benchmark (K=12 on RMBench, K=2 elsewhere), so treating the memory module as plug-and-play across new tasks is optimistic. And the real-world validation, on a Franka Research 3 and a Dobot CR5A with 10 demonstrations per task, is a careful lab protocol with scenes manually reproduced for fair comparison, not a stress test of open-world messiness. What the paper does not give you is a comparison against closed commercial VLAs, or a long-horizon failure profile on truly novel objects.

The direction that is worth watching is smaller teams treating a pre-trained VLM as a fixed substrate and iterating on the memory, action-head, and grounding pieces around it. If that pattern holds, the barrier to a competitive manipulation stack drops meaningfully.