The problem: a photo has no "now"
Ask yourself: staring at a kitchen photo for ten seconds, can you tell whether the kettle water is currently sloshing, or whether a lid is being lifted versus just set down? You can't — a still image has no "in progress," only "is." Yet most VLA models live on exactly that photo: single-frame input, look-act-look. Even Physical Intelligence's flagship π0.5 follows this single-frame paradigm.
StreamPI (arXiv 2608.26067), posted August 26, aims to fix this: upgrade robots from looking at photos to watching video.
How it works
The method is simple in essence:
- Each timestep is packed as a "visual observation + language instruction" temporal unit. Bidirectional attention handles image-text fusion inside a unit; causal attention across units ensures the model only sees the past, never the future.
- The instruction is re-bound at every timestep — a "semantic anchor" so the robot doesn't forget its goal mid-task.
- History frames are not recomputed; they are rolled into a KV cache and reused.
- Zero new parameters. No video encoder, no memory module — just surgery on π0.5's attention structure.
- LIBERO: π0.5 averages 96.9; StreamPI (5-frame) hits 98.3. On the memory-heavy LIBERO-Long: 92.4 → 95.0.
- CALVIN (long-horizon chains): average chain length 4.547 vs π0.5's 4.313 and MemoryVLA's 4.090. By the fifth step in a chain, MemoryVLA drops to 69.4% while StreamPI keeps 85.0% — the memory specialist beaten on memory, attributed to error accumulation over long horizons.
- Real-robot, four tasks (success rate, π0.5 vs StreamPI): Shell Game 46.7% → 80.0%; rolling-object grasping 26.7% → 63.3%; narrow-mouth pen insertion 40% → 66.7%; paper-cup bagging 60% → 92%. All these tasks require knowing "what just happened."
- Cost: inference latency on a 4090 goes from 94.4 ms to 103.6 ms — only 9.2 ms for five frames, thanks to cache reuse.
Training tricks: random temporal intervals between frames and random dropout of early frames, so the model is robust to frame jitter or drops on real hardware. Trained with T=5 frames but fed only 1 frame at inference, it still holds 97.1 average on LIBERO.
The numbers
Caveats
1. Flashy terms like "continuous physical intelligence" and "spatiotemporal intelligence" come only from the press release; the paper itself calls it streaming multimodal temporal modeling. 2. The press release's "96.4% → 97.5%" baseline comes from an ablation table, not π0.5's official 96.9 — easy to under-read the starting point. 3. Real-robot evaluations used only 15–30 trials per task with no error bars; a single run is worth 3–7 percentage points, so the +33.3 on Shell Game may not always replicate.
Disambiguation note
Chinese reports mention "大晓机器人," whose English registered name is ACE ROBOTICS, a Shanghai company chaired by SenseTime co-founder Wang Xiaogang, with Dacheng Tao as chief scientist; Ant Group led its angel round in February — as an investor, not an owner. Ant's own embodied-AI company is Robbyant (蚂蚁灵波), which makes LingBot-VLA — a different company and a different model. StreamPI's technical foundation is Physical Intelligence's open-source openpi, a lineage Chinese coverage omitted entirely.
Eight of the paper's ten authors are from the University of Hong Kong (corresponding author Hengshuang Zhao, SAIL lab). Code is open-sourced at hku-sail/StreamPI (146 stars in four days); the project page promised a public release on August 30.
Embodied AI has spent two years racing on parameters, data, and control frequency while leaving the time dimension empty. StreamPI filled it for 9.2 ms. Whether it's worth it — let the next round of replications decide.
References: Paper arXiv:2608.26067 | Code hku-sail/StreamPI | Project page | Leiphone 8-28 | QbitAI: Ant angel round