Cross-Paper Review — Real-Time Execution for VLA Policies

Topic survey (updated Aug 2026) · how big policies act smoothly at control rate: chunking, continuation, anytime decoding, async systems. Structure: 📈 trend · ⚖️ approaches · ⚠️ limitations. Companion pages: RTC · Legato · OAT · Fast-in-Slow · AsyncVLA · VLA Attention · Ψ₀.

1. The problem

A 2–7B VLA takes 100–300 ms per forward pass; robots need 10–50 Hz control. Action chunking (predict H steps, execute while computing the next chunk) is the universal answer — and its failure mode is the chunk boundary: inference delay plus flow-policy multimodality make consecutive chunks disagree, producing visible hesitation, jitter, and collisions.

2. Trend arc — from test-time patch to trained-in property

Stage	Method	Mechanism	Status
2025	[RTC](/Heungwoo/research/wiki/NeurIPS-2025-Real-Time-Chunking) (test-time)	Inference-time inpainting/guidance constrains the new chunk to overlap the executing one	Widely adopted; external to the policy — spurious multimodal switching, trajectories never intrinsically smooth
Late 2025	Training-time RTC (PI)	Simulate inference delay during training: expose first d ~ U(0, d_max) actions un-noised, mask from loss	Adopted by [Ψ₀](/Heungwoo/research/wiki/Review-Psi0) after test-time guidance proved unstable on their model — a documented negative result for test-time steering
RSS 2026	[Legato](/Heungwoo/research/wiki/RSS-2026-Legato) (native continuation)	Denoising initialized from a schedule-shaped mixture of committed actions + noise; flow dynamics reshaped for train/inference consistency; randomized schedules → controllable smoothness, variable delays	Beats RTC by ≈10% on both smoothness (NSPARC) and completion time across five real tasks
RSS 2026	[OAT](/Heungwoo/research/wiki/RSS-2026-OAT) (anytime decoding)	Ordered token space → every prefix decodes to a valid coarse action; more tokens refine it	A compute-quality dial rather than a boundary fix; AR-policies only

ICML 2026 — the efficiency cluster (9+ papers), strongest of any venue. Streaming: Reflex achieves 50 Hz stable streaming (2.58× speedup, −54% reaction latency) via timestep-invariant attention partitioning with O(1) cache updates. Fewer/faster denoise steps: OMP one-step MeanFlow, STEP warm-started actions (+21.6% over baselines at 2 steps), Sparse ActionGen (4×). Token/channel pruning: GridS (−76% FLOPs, no drop), SpecPrune-VLA (1.57–1.70×), EcoVLA. Reasoning latency: Latent Reasoning VLA internalizes CoT (−90% inference latency), AVA-VLA adds confidence-gated early exit. Policy-agnostic: Speedup Patch (1.8× via safe chunk downsampling). Chunk coherence gets its own treatments (FocalPolicy frequency-optimized chunking).

Parallel line — architectural asynchrony: dual-rate systems (Fast-in-Slow: slow VLM cadence + fast action decoding), AsyncVLA, persistent-context AR experts (AR-VLA, RSS #85: the expert keeps its own history, VL prefixes refresh out-of-band), and action-to-action flow (RSS #209: warm-start denoising from the previous action instead of Gaussian noise).

3. Approach comparison

Approach	Pros	Cons	Choose when
Test-time RTC	Drop-in, no retraining	Multimodal switching; can be unstable (Ψ₀'s experience)	Frozen checkpoint you can't retrain
Training-time continuation (Legato-class)	Intrinsically smooth; delay-robust by randomization; faster completion	Requires retraining; delay range fixed at training	You own the training loop (the 2026 default)
Anytime tokens (OAT)	Graceful degradation under compute pressure	AR-family only	Variable compute budget / early-commit control
Async dual-rate	Decouples semantic and control rates	System complexity; two-model coherence	Reasoning-heavy tasks with fast reflex needs
Single-step action head (IROS 2026 🆕)	[IMLE-VLA](/Heungwoo/research/wiki/IROS-2026-IMLE-VLA) replaces the iterative diffusion/flow head with a 1-step cIMLE generator — 55 Hz (3.67× π0.5), LIBERO 98.0%, keeps multimodality + robustness	Head-swap needs retraining; single-step on high-precision contact untested	You want to kill the multi-step sampling latency at the source
Remote/cloud + RTC	Big models off-robot	Network variance (Qwen-RobotNav: server 196 ms avg but spiky vs Jetson-Thor 204 ms stable)	Fleet ops with reliable links; edge for latency-critical tasks

4. Interactions that bite

• History conditioning raises the denoising bill: RobotManip's in-context variant needs 10 steps where the base needs 4 — robustness features and latency budgets trade off directly.
• Chunk length H couples everything: smoothness machinery, RL credit assignment (RECAP's N-step advantages), and reactivity limits (RobotManip lists fixed chunk length as a stated limitation for sub-second reactive control).
• Wiring choice sets the floor: prefix-KV reuse (π) vs full re-attention per denoise step (concatenation) — argued but never measured head-to-head (Review-VLM-Action-Connection).

5. ⚠️ Limitations

• Latency is the least-reported number in the field. Among 2026 flagships, only scattered figures exist (π0.6 63 ms/chunk-class reports; Ψ₀ ~160 ms/pass; Qwen-RobotNav's deployment table); the Qwen manipulation suite reports none. No standard metric (ms/chunk? effective Hz? jerk?) exists.
• Edge deployment measurement is finally starting: the XPU characterization study profiles VLAs across GPUs/NPUs (compute-bound VLM phase, memory-bound expert phase; 2.9–3.3× via phase-aware scheduling) — but manipulation-side FP8/TensorRT deployment reports remain rare.
• Smoothness metrics (NSPARC etc.) are young; no benchmark scores hesitation/recovery under forced delay spikes.
• All continuation methods assume the policy is the bottleneck; perception-latency (multi-view encoding) is unaddressed.

← Back to Home · Reviews