Action Chunks Skrrt Skrrt

TL;DR

Second article on file from Binh Pham’s personal wiki (garden.binhph.am), landing two days after his Closed-Loop Policies Closed-Loop Policies note. Per Sid’s Slack framing the post is a “timely” write-up on speeding up robot movement during inference — the same design axis as the async-inference / chunk-boundary / flow-matching-latency thread on the wiki (πR², REMAC, A2A, Legato, FASTER, FlashVLA, VLA-RAIL, TurboVLA, Cosmos 3 Edge, Embodied.cpp, APXinf-robo, C³ache). The blog itself was unreachable at filing (HTTP 429 across ~10 retry attempts spanning the ingest window), so this page is filed from Sid’s Slack note plus context from the author’s prior filed work and the surrounding literature. The specific lever the article uses to speed up robot motion at inference is not recoverable at filing time.

Key claims

• Post concerns speeding up robot movement during inference [per Sid’s Slack note, 2026-09-18].
• Second entry in Binh Pham’s garden.binhph.am series, following Closed-Loop Policies on 2026-09-16.

Method

Article content was not retrievable at filing time — garden.binhph.am returned HTTP 429 across ~10 successive fetch attempts spanning the ingest window, matching the same rate-limit behavior seen when the sibling Closed-Loop Policies Closed-Loop Policies article was filed two days earlier. This page is therefore filed from Sid’s one-line summary (“timely post on speeding up robot movement during inference”) plus context from the author’s prior filed work.

Author is Binh Pham (pham-tuan-binh on GitHub, garden.binhph.am blog, build_with_binh on YouTube), an independent hobbyist / content creator with two prior filed artifacts: ACT in the browser — three.js SO-101 + leader-arm teleop + trained model running locally on the web — an in-browser lerobot ACT training + inference demo at lerobot.binhph.am — and Closed-Loop Policies — a further data point on the action-history × execution-horizon design space aligned with Revisiting Open-Loop Execution in Robotics: Toward Reactive, Higher-Performing Policies Zeng et al., Simulation Pre-training for Dexterity (SPD) SPD, and Behavioral Cloning Mystery Seohong Park. The through-line of Pham’s public work is careful hands-on characterization of the deployment-time properties of behavior-cloning VLA policies (browser-side ACT, action-history / execution-horizon empirics, and — inferentially, from the title and Slack framing — chunk-inference latency).

Exact methodology, tasks, base policy, hardware, quantitative numbers, and the specific latency lever discussed are not recoverable at filing time.

Results

Not recoverable at filing time (HTTP 429). Will need a /bud refresh once garden.binhph.am is reachable.

Why it’s interesting

This is the third filed contribution from an independent hobbyist pushing empirical characterization of production-relevant deployment-time properties of BC / VLA policies — a category otherwise dominated by well-resourced industrial labs on this wiki. Two prior instances from the same author (ACT in the browser — three.js SO-101 + leader-arm teleop + trained model running locally on the web in-browser ACT loop; Closed-Loop Policies action-history / execution-horizon empirics) both landed on axes where the industrial labs’ claims turned out to be reproducible and generalizable — the value of a fresh independent replication signal here is exactly that.

The “speeding up robot movement during inference” framing sits squarely on the wiki’s growing async-inference / chunk-latency board. On the training-time side the field has Real-Time Robot Execution with Masked Action Chunking REMAC (LoRA + prefix masking + self-conditioned curriculum), πR²: Reactive Real-time Flow Policies πR² (slow/fast channel disentanglement + latency-adaptive noise staircase), A2A: Action-to-Action Flow Matching A2A (anchor flow endpoint to previous proprioceptive action for 1-step generation), Legato: Learning Native Continuation for Action Chunking Flow Policies Legato (native chunk-continuation fine-tuning). On the inference-runtime side there is Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots Embodied.cpp (llama.cpp-shaped runtime), APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin APXinf-robo (Jetson Thor FP8/INT8 with 41 ms / 24 Hz for π0.5), C³ache: Accelerating World Action Models with Cross Inference Chunk Cache C³ache (cross-chunk DiT residual caching), and TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM TurboVLA (sub-billion-parameter LLM-free action head at 31 ms on RTX 4090). Which of these levers Pham’s article uses (or whether it stakes out a fifth) is exactly what the 429 is preventing us from seeing.

See also

• Closed-Loop Policies — sibling article from the same author, filed two days earlier; same rate-limit issue, filed from Slack summary
• ACT in the browser — three.js SO-101 + leader-arm teleop + trained model running locally on the web — same author’s earlier filed work: in-browser lerobot ACT training + inference
• VLA Models — recipe-lever board; this article sits on the inference-latency axis
• LLM Inference Efficiency — the concept page tracks the analogous problem for LLMs; robot-policy inference has now imported multiple LLM-side ideas (KV caching, speculative decoding, quantization)
• πR²: Reactive Real-time Flow Policies — architectural approach to sub-100 ms flow-matching VLA inference
• Real-Time Robot Execution with Masked Action Chunking — training-time recipe for async-inference robustness (REMAC)
• Legato: Learning Native Continuation for Action Chunking Flow Policies — native chunk-continuation fine-tuning for faster reactive control
• Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots — the deployment-runtime substrate side of the same problem