# StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models

Zhe Liu1*    Jinghua Hou1*    Yuxiang Lu1    Zhenya Yang1    Xianzhe Fan1    Junwei Luo1
Junyi Li1    Ruihua Han1    Zhi Hou2    Hengshuang Zhao1†

1 The University of Hong Kong    2 ACE Robotics
* Equal contribution    † Corresponding author

arXiv: 2608.26067 Project Page Demo Video

Official implementation of StreamPI, built on openpi, with additional, fully validated support for JAX multi-node distributed training beyond the standard single-node setup, enabling substantially more efficient training without compromising model performance.
## TL;DR Single-frame Vision-Language-Action (VLA) models such as $\pi_{0.5}$ cannot retain past observations, while window-based multi-frame models repeatedly process the full history at substantial computational cost. **StreamPI** instead treats every (visual observation, language instruction) pair as an atomic temporal unit. Bidirectional attention within each pair preserves cross-modal grounding, causal attention across pairs captures temporal context, and a KV cache avoids recomputing past observations. This equips a pretrained single-frame VLA with temporal reasoning without adding parameters, while supporting flexible context lengths and efficient streaming inference.

Overview of the StreamPI architecture and streaming inference workflow

## 🔥 Highlights - **Persistent temporal reasoning:** StreamPI equips $\pi_{0.5}$ with memory and multi-frame geometric cues while preserving its pretrained representation space. - **Instruction-anchored modeling:** Repeating the instruction in every temporal unit prevents the task goal from being diluted as visual context grows. - **No additional parameters:** Temporal modeling is implemented entirely through an extended token sequence and a block-wise attention mask. - **Asynchrony-aware training:** Random-interval sampling and temporal masking expose the policy to variable observation timing and incremental context. - **Efficient deployment:** On a single RTX 4090, increasing the context from one to five frames adds only 9.2 ms of mean inference latency (94.4 ms to 103.6 ms). ## 🎥 Demo

StreamPI real-world demonstration preview
▶ Click to play the 2:07 real-world demonstration (MP4)

The animated preview showcases StreamPI on precise perception and memory-dependent manipulation tasks. Click it to play the full video. ## 📰 News - **August 30, 2026:** Code and model weights will be released. - **August 26, 2026:** The paper is released on arXiv. ## ✨ Abstract Vision-Language-Action models have demonstrated strong performance in robot manipulation, yet leading models such as $\pi_{0.5}$ still operate one frame at a time. This single-frame paradigm limits both temporal memory and the precise spatial perception that can emerge from aggregating observations over time. **StreamPI** is a streaming multimodal temporal modeling framework that adds these capabilities without introducing any new model parameters. The central design is *instruction-anchored temporal modeling*. StreamPI treats each (visual observation, language instruction) pair as an atomic temporal unit: bidirectional attention within a pair enables full cross-modal fusion, while causal attention across pairs supports autoregressive streaming inference. The instruction therefore remains a persistent semantic anchor throughout execution. A complementary *random-interval streaming training* strategy improves robustness to variable frame rates and asynchronous observation arrival. By relying on the LLM backbone's length extrapolation, StreamPI inherits all pretrained $\pi_{0.5}$ weights and supports both single-frame and multi-frame inference. Experiments on real robots, LIBERO, and CALVIN show consistent gains from temporal modeling. ## 🧠 Method ### Instruction-Anchored Temporal Modeling At time $t$, the base VLA receives multi-view visual observations $\mathbf{V}_t$ and a language instruction $l_t$. StreamPI binds them into an indivisible temporal unit and concatenates the latest $T$ units: $$ \mathbf{u}_t = (\mathbf{V}_t, l_t), \qquad \mathbf{U} = [\mathbf{u}_{t-T+1}, \ldots, \mathbf{u}_t]. $$ The attention pattern operates at two levels: - **Intra-pair bidirectional attention** allows all visual and language tokens within $\mathbf{u}_\tau$ to interact, producing a semantically grounded representation $\mathbf{h}_\tau$. - **Inter-pair causal attention** allows the current representation to attend to earlier temporal units, but never to future ones. Re-anchoring the instruction at every time step prevents instruction forgetting as the temporal horizon grows. The entire hierarchy is realized by restructuring the attention mask and extending the input token sequence, so StreamPI inherits all pretrained $\pi_{0.5}$ weights without architectural changes or additional parameters. ### Random-Interval Streaming Training Real robots produce asynchronous observation streams with variable timing. At each historical sampling step $i$, StreamPI perturbs a base interval $\bar{\delta}$ with $\epsilon_i \sim \mathcal{U}(-\Delta, +\Delta)$ and clips $\delta_i = \bar{\delta} + \epsilon_i$ to $[\delta_{\min}, \delta_{\max}]$. Historical frames are selected at cumulative offsets such as $t-\delta_1$ and $t-\delta_1-\delta_2$. In the paper's training protocol, each interval is sampled uniformly from $[3, 7]$. StreamPI also samples a masking count $k \in \{0, \ldots, T-1\}$ and hides the earliest $k$ units in a training sequence. This temporal masking simulates the incremental context available at the beginning of streaming inference. Together, randomized intervals and temporal masking reduce the mismatch between synchronous training data and asynchronous deployment. ### Streaming Inference At the first time step, StreamPI encodes the current temporal unit, predicts an action chunk, and stores the resulting Key and Value representations. Each subsequent call encodes only the newly arrived unit, which attends to the cached history. The cache is bounded by the configured context length $T$ and is flushed when the next unit would exceed that limit. This eliminates redundant re-encoding of past observations while retaining the temporal evidence needed for long-horizon manipulation. ## 📊 Experiments ### Real-Robot Manipulation The real-robot platform uses AgileX PiperX 6-DoF arms in an ALOHA-style leader-follower setup, with one front-view Intel RealSense D455 and two wrist-mounted RealSense D435 cameras. For each task, we collect 100 teleoperated demonstrations at 30 FPS and fully fine-tune the pretrained $\pi_{0.5}$ policy. Evaluation covers two complementary task categories: - **Precise perception:** Cup Insertion into Cup Sleeve and Pen Insertion into Narrow Bottle. - **Memory dependence:** Rolling Object Grasping and Shell Game. StreamPI improves over the single-frame baseline on every task, with gains ranging from 26.7 to 36.6 percentage points. | Task | Trials | $\pi_{0.5}$ | StreamPI | Gain | | -------------------------------- | -----: | -----------: | -------: | ---: | | Cup Insertion into Cup Sleeve | 25 | 60.0% | **92.0%** | +32.0 pp | | Pen Insertion into Narrow Bottle | 30 | 40.0% | **66.7%** | +26.7 pp | | Rolling Object Grasping | 30 | 26.7% | **63.3%** | +36.6 pp | | Shell Game | 15 | 46.7% | **80.0%** | +33.3 pp | ### LIBERO Simulation Benchmark LIBERO contains four suites - Spatial, Object, Goal, and Long - with 10 tasks per suite and 50 trials per task. Success rates (%) are shown below. StreamPI ($T=5$) reaches **98.3% average success**, improving over $\pi_{0.5}$ by 1.4 points overall, 2.8 points on LIBERO-Goal, and 2.6 points on LIBERO-Long. Performance on LIBERO-Spatial remains tied at 98.8%, consistent with that suite's emphasis on static spatial relations already visible in one frame. | Method | Spatial | Object | Goal | Long | Avg. | | --------------------- | ------: | -----: | ---: | ---: | ---: | | Diffusion Policy | 78.3 | 92.5 | 68.3 | 50.5 | 72.4 | | Octo | 78.9 | 85.7 | 84.6 | 51.1 | 75.1 | | SpatialVLA | 88.2 | 89.9 | 78.6 | 55.5 | 71.7 | | TraceVLA | 84.6 | 85.2 | 75.1 | 54.1 | 74.8 | | OpenVLA | 84.7 | 88.4 | 79.2 | 53.7 | 75.9 | | CoT-VLA | 87.5 | 91.6 | 87.6 | 69.0 | 81.1 | | $\pi_0$-FAST* | 96.4 | 96.8 | 88.6 | 60.2 | 85.0 | | SmolVLA | 93.0 | 94.0 | 91.0 | 77.0 | 88.8 | | GR00T-N1 | 94.4 | 97.6 | 93.0 | 90.6 | 93.9 | | UniVLA | 95.4 | 98.8 | 93.6 | 94.0 | 95.4 | | FLOWER | 97.1 | 96.7 | 95.6 | 93.5 | 95.7 | | CronusVLA | 90.1 | 94.7 | 91.3 | 68.7 | 86.2 | | TriVLA | 91.2 | 93.8 | 89.8 | 73.2 | 87.0 | | 4D-VLA | 93.8 | 92.8 | 95.6 | 86.5 | 92.2 | | CogACT | 87.5 | 90.2 | 80.2 | 53.2 | 77.8 | | ST-$\pi$ | 98.4 | 98.3 | 96.9 | 94.3 | 97.3 | | MemoryVLA | 98.4 | 98.4 | 96.4 | 93.4 | 96.5 | | $\pi_0$ | 96.8 | 98.8 | 95.8 | 85.2 | 94.2 | | $\pi_{0.5}$ | **98.8** | 98.2 | 96.8 | 92.4 | 96.9 | | **StreamPI ($T=3$)** | **98.6** | 98.6 | 98.6 | 93.8 | 97.5 | | **StreamPI ($T=5$)** | **98.8** | **99.8** | **99.6** | **95.0** | **98.3** | ### Key Ablations Our ablations isolate the contribution of each temporal design choice: | Component | Strong setting | Comparison | Avg. gain (pp) | LIBERO-Long gain (pp) | | --------- | -------------- | ---------- | --------: | ---------------: | | Intra-pair fusion | Bidirectional attention, $T=5$ | Causal intra-pair attention | +2.8 | +4.4 | | Inter-pair context | Five frames with causal inter-pair attention | Single-frame context | +1.8 | +3.0 | | Temporal sampling | Random $\delta \sim \mathcal{U}[3,7]$, $T=5$ | Fixed $\delta=1$ | +1.3 | +1.6 | A model trained with $T=5$ also generalizes to shorter inference contexts: it obtains 97.4% average success at test-time $T=3$ and 97.1% at $T=1$, both above the 96.5% single-frame ablation baseline. ### Streaming Inference Efficiency Mean latency over 20 real-robot trials on one NVIDIA GeForce RTX 4090. Extending the temporal context from one to five frames adds only 9.2 ms. | Streaming frames | Inference time (ms) | Overhead vs. $T=1$ | | ---------------: | ------------------: | ------------------: | | 1 | $94.4 \pm 3.4$ | N/A | | 3 | $97.9 \pm 5.1$ | 3.5 ms | | 5 | $103.6 \pm 6.3$ | 9.2 ms | | 8 | $110.9 \pm 10.2$ | 16.5 ms | | 10 | $117.9 \pm 16.5$ | 23.5 ms | ### CALVIN Benchmark The StreamPI paper also evaluates long-horizon temporal reasoning on CALVIN, where a policy must execute sequences of up to five consecutive tasks. Success rates (%) are reported at each sequence position; Avg. is the average number of consecutively completed tasks. | Method | 1 | 2 | 3 | 4 | 5 | Avg. | | --------------------- | ---: | ---: | ---: | ---: | ---: | ---: | | MemoryVLA | 94.8 | 87.4 | 81.4 | 75.9 | 69.4 | 4.090 | | $\pi_{0.5}$ | 94.2 | 88.7 | 85.7 | 83.2 | 79.5 | 4.313 | | **StreamPI ($T=5$)** | **96.9** | **93.6** | **90.7** | **88.5** | **85.0** | **4.547** | ## ⚙️ Setup ### Requirements StreamPI follows the software requirements of [openpi](https://github.com/Physical-Intelligence/openpi) and requires Python 3.11 or later. Because StreamPI stacks $T$ temporal units in the prefix, training memory grows with the temporal horizon. Model parallelism can be configured with `--fsdp-devices`; multi-node JAX training is provided by [`scripts/train_multi_node.py`](scripts/train_multi_node.py). The reference experiments in the paper use the following compute settings: | Experiment | Hardware | Batch size | Training steps / trials | | ---------- | -------- | ---------: | ----------------------: | | LIBERO training | 8 x NVIDIA H100 | 256 | 30,000 steps | | Real-robot training | 8 x NVIDIA H100 | 128 | 50,000 steps | | Latency evaluation | 1 x NVIDIA RTX 4090 | N/A | 20 trials | ### Installation Clone the repository with its submodules, then install the dependencies with [uv](https://docs.astral.sh/uv/): ```bash git clone --recurse-submodules https://github.com/happinesslz/StreamPI.git cd StreamPI GIT_LFS_SKIP_SMUDGE=1 uv sync GIT_LFS_SKIP_SMUDGE=1 uv pip install -e . ``` If the repository was cloned without submodules, initialize them separately: ```bash git submodule update --init --recursive ``` Use `uv run ` for the commands below, or activate the generated environment once with: ```bash source .venv/bin/activate ``` For a container-based setup, see [`docs/docker.md`](docs/docker.md). ## 🚀 Usage ### Training Configurations StreamPI configurations are defined in [`src/openpi/training/config.py`](src/openpi/training/config.py). The temporal context length $T$ is controlled by `Pi0Config.hist_horizon`; frame spacing and temporal jitter are controlled by `DataConfig.hist_interval`, `jitter_range`, and `enable_jitter`. | Configuration | Benchmark | $T$ | Base interval | Jitter enabled by default | | ----------------------- | --------- | --: | ------------: | :-----------------------: | | `pi05_libero_stream3` | LIBERO | 3 | 5 | No | | `pi05_libero_stream5` | LIBERO | 5 | 5 | No | | `pi05_calvin_stream3` | CALVIN | 3 | 5 | No | | `pi05_calvin_stream5` | CALVIN | 5 | 5 | No | All streaming configurations fine-tune from the pretrained single-frame checkpoint at `gs://openpi-assets/checkpoints/pi05_base/params`. Because StreamPI introduces no new parameters, it can inherit the pretrained $\pi_{0.5}$ weights directly. > [!IMPORTANT] > The results reported in the paper use independently randomized intervals $\delta \sim \mathcal{U}[3,7]$. The checked-in LIBERO and CALVIN configurations currently set `enable_jitter=False`, while [`TemporalJitter`](src/openpi/transforms.py) applies one shared frame offset around a fixed interval. Exact reproduction of the paper's random-interval experiments therefore requires aligning both the configuration and data-sampling path with the paper protocol. ### 1. Prepare Normalization Statistics Precomputed statistics for LIBERO and CALVIN are included under `assets/`. To recompute them: ```bash # LIBERO bash compute_norm.sh # CALVIN (writes to assets/pi05_calvin/.../norm_stats.json) python compute_norm_calvin.py ``` ### 2. Launch Training For the reported results, we fully fine-tune the pretrained $\pi_{0.5}$ weights using the same optimizer and learning-rate schedule as the base model. A single-node LIBERO run can be launched with: ```bash JAX_DEBUG_NOWAIT=1 XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 \ python scripts/train.py pi05_libero_stream5 \ --exp-name=pi05_stream5_libero \ --overwrite \ --batch-size=256 \ --fsdp-devices=8 ``` For the repository's CALVIN training examples, see [`train_stream5_calvin.sh`](train_stream5_calvin.sh) and [`train_stream3_calvin.sh`](train_stream3_calvin.sh). #### JAX Multi-Node Distributed Training StreamPI extends the upstream `openpi` training pipeline with a **fully validated JAX multi-node distributed training path**. [`scripts/train_multi_node.py`](scripts/train_multi_node.py) uses `jax.distributed.initialize` to form one global device mesh across all nodes, combines data parallelism with Fully Sharded Data Parallel (FSDP), partitions the global batch across processes, and coordinates checkpoint saving and restoration across hosts. The model architecture, training objective, optimizer, and learning-rate schedule remain unchanged, so multi-node training preserves single-node model performance while **substantially reducing wall-clock training time**. Run the same training command on every node after setting the shared coordinator address and a unique zero-based rank for each node: ```bash export MASTER_ADDR= export MASTER_PORT=6060 export NNODES=4 export NODE_RANK=<0-3> ``` `--batch-size` denotes the global batch size and must be divisible by the global JAX device count and the number of nodes. `--fsdp-devices` controls the width of the FSDP axis and must divide the total number of devices. For example, four 8-GPU nodes with `--fsdp-devices=8` form a global mesh with four-way data parallelism and eight-way FSDP. Ready-to-use four-node launchers are provided in [`train_stream5_calvin_4_nodes.sh`](train_stream5_calvin_4_nodes.sh) and [`train_stream5_libero_4_nodes.sh`](train_stream5_libero_4_nodes.sh). A LIBERO launch command is: ```bash python -u scripts/train_multi_node.py pi05_libero_stream5 \ --exp-name=pi05_stream5_libero_4nodes \ --overwrite \ --batch-size=256 \ --fsdp-devices=8 ``` Checkpoints are written to: ```text checkpoints/// ``` ### 3. Evaluation #### LIBERO See [`examples/libero/README.md`](examples/libero/README.md) for environment setup and evaluation options. To evaluate a custom checkpoint, run the LIBERO client and policy server in separate processes: ```bash # Client python examples/libero/main.py # Policy server uv run scripts/serve_policy.py --env=LIBERO policy:checkpoint \ --policy.config=pi05_libero_stream5 \ --policy.dir=/path/to/checkpoint ``` #### CALVIN See [`examples/calvin/README.md`](examples/calvin/README.md) for the complete setup. Start the policy server in the StreamPI environment: ```bash bash run_eval_calvin_benchmark.sh # Equivalent command: python scripts/serve_policy.py --port=8000 policy:checkpoint \ --policy.config=pi05_calvin_stream5 \ --policy.dir=checkpoints/pi05_calvin_stream5/pi05_stream5_calvin_4nodes/29999 ``` Then run the client in a separate CALVIN environment: ```bash export CALVIN_ROOT=/path/to/calvin bash run_calvin_env.sh ``` Evaluation outputs are saved to: ```text //result.json //success_rate.txt ``` ### Streaming Inference and Deployment Streaming KV-cache inference requires no additional server flags. Any configuration with `hist_horizon > 1` uses the temporal memory path in `Pi0.sample_actions`. The client controls cache boundaries through the `step` field included in each observation: 1. Send a `step` counter with every inference request. 2. The server starts a new cache whenever `step % hist_horizon == 0`. 3. Between resets, only the newly arrived temporal unit is encoded and allowed to attend to the cached history. In the paper, simulation uses a fixed inference interval of $\delta=5$, while real-robot deployment samples $\delta \sim \mathcal{U}[3,7]$. This cadence is controlled by the client and determines how many environment frames or actions elapse between policy calls. The CALVIN client in [`examples/calvin/main.py`](examples/calvin/main.py) implements the counter. Set `--replan_steps` to match `hist_interval` (5 by default). The current LIBERO client sends a constant `step: 0`, so it uses full re-encoding rather than the KV-cache path; increment `step` to exercise streaming inference on LIBERO. For real-robot deployment, start a policy server with [`scripts/serve_policy.py`](scripts/serve_policy.py), connect the robot client over WebSocket through [`packages/openpi-client`](packages/openpi-client), and include the instruction and `step` counter in each request. The `pi05_stream5_aloha_shell_game` configuration illustrates the data transforms used for the AgileX setup. ## ⚠️ Limitations - Training still loads all $T$ frames jointly, so the cost becomes prohibitive for extremely long temporal horizons. - Random-interval training improves robustness to timing variation but does not fully address extreme real-world asynchrony. - Promising future directions include efficient training beyond 100 frames and adaptive KV-cache pruning. ## 📖 Citation If you find StreamPI useful in your research, please cite: ```bibtex @article{liu2026streampi, title = {StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models}, author = {Liu, Zhe and Hou, Jinghua and Lu, Yuxiang and Yang, Zhenya and Fan, Xianzhe and Luo, Junwei and Li, Junyi and Han, Ruihua and Hou, Zhi and Zhao, Hengshuang}, journal = {arXiv preprint arXiv:2608.26067}, eprint = {2608.26067}, archivePrefix = {arXiv}, primaryClass = {cs.CV}, url = {https://arxiv.org/abs/2608.26067}, year = {2026} } ``` ## 🙏 Acknowledgements StreamPI is built on [openpi](https://github.com/Physical-Intelligence/openpi). We also thank the authors of [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) and [CALVIN](https://github.com/mees/calvin) for their open-source benchmarks and codebases. ## 📄 License This project is released under the terms described in [`LICENSE`](LICENSE). The Gemma model components are subject to the terms in [`LICENSE_GEMMA.txt`](LICENSE_GEMMA.txt).