FASTER¶

Development line: project:faster · thread faster
Last event: 2026-03-20 · 1 dated since 2026-03-20 · Researched: 2026-09-05 · confidence: medium

What it is¶

FASTER is an open-source fast-action sampling and streaming deployment method for teams fine-tuning flow Vision-Language-Action policies, not a standalone robot policy.

• Horizon-Aware Schedule prioritizes near-term actions.
• Streaming client-server path dispatches early actions while later actions refine.
• Current code builds on JAX openpi and documents π0.5 and AgileX Piper workflows.

Development line¶

• 2026-03-20 — FASTER public website and repository were linked. The public repository recorded two README updates to document the project and paper state.

What changed¶

2026-03-19: The FASTER paper entered arXiv as v1. 2026-03-20: The public repository recorded two README updates to document the project and paper state. 2026-04-29: arXiv v2 appeared without a technical changelog. 2026-04-30: FASTER v1.0 open-sourced the official implementation. 2026-05-16: arXiv v3 became the current paper version without a revision note.

How to use this¶

As of 2026-03-20, use the FASTER public website and GitHub repository for project documentation and source access; we have verified no further usage or release claim.

1. Start from a JAX openpi flow-policy workflow and an already mapped robot platform; use pi05_faster_agilex as the maintained reference rather than treating FASTER as a pretrained standalone policy. — https://github.com/innovator-zero/FASTER
2. Clone with submodules, then create the documented uv environment using GIT_LFS_SKIP_SMUDGE=1 uv sync and GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .. — https://github.com/innovator-zero/FASTER
3. Prepare a LeRobot v2.1 dataset and platform input/output mapping, then compute normalization statistics for the FASTER config. — https://github.com/innovator-zero/FASTER
4. Fine-tune π0.5 with pi05_faster_agilex and point the policy server at the resulting checkpoint. — https://github.com/innovator-zero/FASTER
5. Run the policy server and robot-controller client over WebSocket; select asynchronous HAS or streaming inference according to the task’s reaction-time requirement. — https://github.com/innovator-zero/FASTER

Best practices¶

• Learn the original openpi workflow before adapting FASTER. — https://github.com/innovator-zero/FASTER
• Keep HAS inference settings alpha and u0 matched to the values used during training. — https://github.com/innovator-zero/FASTER
• Set controller delay one step above measured inference delay to cover runtime and network variation, and keep execution horizon at least as large as delay. — https://github.com/innovator-zero/FASTER
• Use wired LAN for remote inference to reduce latency and packet loss. — https://github.com/innovator-zero/FASTER
• Reserve streaming with a small execution horizon for highly dynamic tasks; on ordinary pick-and-place or folding tasks it can add motion jitter. — https://github.com/innovator-zero/FASTER
• Respect the documented memory limits and the current boundary: full π0.5 fine-tuning was tested on Ubuntu 22.04, and multi-node training is not supported. — https://github.com/innovator-zero/FASTER

Superseded by this¶

• 2026-04-30: the v1.0 open-source release superseded the earlier state where FASTER had no public implementation.
• 2026-05-16: arXiv v3 superseded arXiv v1 as the current paper text.

Still unknown¶

• Official materials provide no checkpoint download and no complete reproduction recipe for the reported real-robot table-tennis setup; documentation assumes locally trained checkpoints.
• arXiv timestamps v2 and v3 without a technical changelog, so their exact deltas from v1 remain unknown.
• We found no first-party Simplified-Chinese technical documentation; available Chinese explanations are secondary.

Sources¶

Agent brief¶

• Subject: project:faster, thread faster, 1 dated events 2026-03-20 → 2026-03-20.
• Practical note: As of 2026-03-20, use the FASTER public website and GitHub repository as starting points for project documentation and source access; we have verified no further usage or release claim.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.