Your robot isn't slow because of bad hardware. It's slow because researchers have been pruning in the wrong place—and nobody noticed.
image from Gemini Imagen 4
Researchers at Bosch Corporate Research have developed ETA-VLA, a vision-language-action model that achieves 61% inference FLOP reduction by aggressively pruning 85% of visual tokens while retaining 94% accuracy on the NAVSIM v2 autonomous driving benchmark. The key innovation is RoPE-free Semantic Scoring, which removes positional encoding to identify decision-relevant tokens based on meaning rather than spatial location, allowing peripheral but critical information like blind spots to be preserved. ETA-VLA scores 85.0 EPDMS, competitive with DiffusionDriveV2 (85.5) but trailing HiST-VLA (88.6) and ELF-VLA (87.1), presenting a compute-accuracy tradeoff for embodied AI applications.
A robot that can see and act is only useful if it can do both fast enough to matter. Right now, most vision-language-action models — the kind of AI that lets a robot look at a scene and decide what to do — run at around five inference cycles per second. That sounds fast. It isn't. A robot operating at 5 Hz in a dynamic environment is the difference between catching a falling cup and watching it shatter on the floor. Closing that gap is the central engineering problem in embodied AI right now. A new paper from researchers at Bosch Corporate Research offers one approach: prune the visual tokens aggressively, and do it inside the language model rather than before it.
The system, called ETA-VLA (Efficient Token Adaptation via Temporal Fusion and Intra-LLM Sparsification), is described in a preprint posted to arXiv on March 26, 2026 (https://arxiv.org/abs/2603.25766). It reduces inference FLOPs by 61 percent while retaining 94 percent of the original accuracy on the NAVSIM v2 autonomous driving benchmark. More striking: it achieves this by cutting 85 percent of visual tokens. The remaining 15 percent carry almost all the decision-relevant information, if you know how to choose them.
That "if" is where the work lives. Standard token pruning approaches score tokens by how far they are from the model's current focus — a positional bias, like always looking at the center of a photograph first. The ETA-VLA team, led by Yiru Wang at Bosch Corporate Research, strips out positional encoding entirely using a technique they call RoPE-free Semantic Scoring. Instead of asking "where is this token?" they ask "what does this token actually mean for the task?" The query and key states are computed without rotational position embedding, removing the distance bias that causes models to underweight peripheral but semantically critical information — a car's blind spot, say, or a pedestrian stepping out from behind a bus.
The mechanism that makes this safe rather than reckless is called diversity-preserving per-view recycling. Globally, the model aggressively prunes task-irrelevant regions. Locally, within each camera view, it enforces a minimum retention floor — a handful of tokens must survive regardless of their semantic score. The paper describes it as mimicking human attention allocation: you focus on the road, but you never completely stop being aware of your peripheral vision. The result is that critical spatial regions like blind spots survive even when the model is running lean.
The numbers sit in an awkward place. ETA-VLA scores 85.0 EPDMS (effective decision-making score) on the NAVSIM v2 Navtest split. The current leader on that benchmark is HiST-VLA at 88.6 EPDMS. ELF-VLA sits at 87.1 — a gap of roughly 2.1 points from ETA-VLA's 85.0. DiffusionDriveV2 achieves 85.5. ETA-VLA is competitive with the latter, clearly behind the former two. Whether that tradeoff is acceptable depends on what you're optimizing for. If you need every percentage point of driving precision, HiST-VLA wins. If you need to run that model on edge hardware at high frequency, ETA-VLA's efficiency lead may matter more than its accuracy deficit.
That is the bet. And it is a bet that matters beyond academic benchmarks.
VLA inference running at under 5 Hz has been a known bottleneck for at least two years. VLA-IAP (https://arxiv.org/abs/2603.22991) notes that processing long visual sequences creates latency that "limits inference to under 5 Hz, far from the required frequency for robust closed-loop robotic control." The robotics field has attacked this from multiple directions simultaneously: VLA-IAP, another token pruning approach, achieves up to 1.54x speedup on the LIBERO manipulation benchmark through training-free pruning (https://arxiv.org/abs/2603.22991). BitVLA goes the quantization route — 1-bit models for edge deployment. OxyGen achieves up to 3.7x speedup through KV cache management. Layer skipping via DySL-VLA is another line of attack. ETA-VLA is the latest entry in what is becoming a crowded field of efficiency plays, each approaching the same wall from a different angle.
The human side of this is straightforward: a robot that reacts faster is safer to be around. A warehouse robot that can adjust mid-reach is less likely to clip a coworker. A humanoid that can process a cluttered kitchen scene and act before a child moves is more useful in a home. Speed enables capability, and the efficiency work being done across VLA compression, quantization, and pruning is what makes speed achievable on real hardware.
No humanoid company has publicly adopted ETA-VLA's approach yet. Figure AI's Helix system, described in a blog post on the company's website (https://www.figure.ai/news/helix), merges visual features from multiple cameras in a multiscale stereo network before tokenization — a pre-LLM compression step rather than the intra-LLM sparsification ETA-VLA uses. The architectural approaches differ, but both teams are trying to solve the same underlying problem: vision sequences generate too many tokens for fast inference. ETA-VLA cuts tokens after the fact. Helix cuts them before they enter the model. Whether one approach proves superior in deployment — or whether they prove complementary — is an open question.
Bosch's position as the source matters here. Corporate research labs publish papers with engineering credibility that academic groups sometimes lack, but they also have commercial incentives that academic researchers don't. The paper notes the work was supported by Bosch Corporate Research. The researchers — Wang, Jiang, S. Wang, Heng, Gu, and Sun — have a concrete application in mind, even if they don't name it. Whether that application is automotive, industrial, or something else isn't specified. The paper is careful not to oversell the humanoid robotics angle, which is either intellectual honesty or a missed opportunity to broaden the audience.
The accuracy gap to HiST-VLA is the most honest concern with ETA-VLA as presented. In a driving context, 3.6 EPDMS points could represent the difference between a system that handles edge cases well and one that doesn't. The 94 percent accuracy retention figure sounds impressive until you remember that 6 percent of something can be the thing that matters most. The diversity-preserving recycling is meant to handle exactly this — to make sure the dropped information isn't critical information — but the mechanism hasn't been tested independently outside the NAVSIM evaluation.
This is worth watching for anyone building or investing in embodied AI. The efficiency race in VLA inference is real, and ETA-VLA adds a genuinely novel mechanism to the toolkit. Whether the semantic pruning inside the LLM proves more effective than pre-tokenization compression (Helix's approach) or other pruning methods (VLA-IAP) will depend on comparative benchmarking the community hasn't done yet. The paper makes a credible case that inside-the-model pruning is the right place to do this work. That case needs to be tested in the wild — on a real robot, in a real environment, next to a real person — before it becomes anything more than an interesting result on a benchmark.
Story entered the newsroom
Research completed — 0 sources registered. ETA-VLA (Bosch, arXiv March 26 2026) does token pruning inside the LLM using RoPE-free semantic scoring — a novel approach that avoids positional bias
Draft (1127 words)
Reporter revised draft based on fact-check feedback
Reporter revised draft based on fact-check feedback (1129 words)
Approved for publication
Headline selected: Your Robot Can't Catch a Falling Cup. This Fixes That.
Published (1133 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Robotics · 13h 23m ago · 4 min read
Robotics · 13h 54m ago · 4 min read
