LATENCY WITHOUT HARDWARE IS MEANINGLESS

Physical AI Inference Latency Leaderboard

Compare published robot-model inference latency with GPU, precision, model size, batch size and test conditions kept visible beside every result.

4 retained recordsLast verified: August 13, 2026Primary-source fields preserved per record

Accelerator

4 of 4 records shownFilters are local and do not create crawlable URLs

TurboVLA

Research result

Model: TurboVLA
Version: 216.1M LIBERO configuration
Hardware: NVIDIA RTX 4090
Precision: Paper configuration; precision should be checked in experiment details
Model size: 216.1M parameters
Batch: 1
Inference latency: 31.2 ms per policy inference
Action frequency: ~32 policy predictions/s
Control frequency note: Full robot servo loop not claimed to run at 32 Hz
Robot / benchmark: Benchmark + AgileX Piper transfer
Test conditions: Observation-to-action-chunk policy latency on one RTX 4090; excludes full sensing/network/servo pipeline
Source: TurboVLA project

TurboVLA

Research result

Model: TurboVLA
Version: 0.4B RoboTwin bimanual configuration
Hardware: NVIDIA RTX 4090
Precision: Paper configuration
Model size: 0.4B parameters
Batch: 1 / paper inference setup
Inference latency: 43.4 ms
Action frequency: ~23 policy predictions/s from latency
Control frequency note: Action chunk = 50 steps; low-level controller frequency is separate
Robot / benchmark: RoboTwin 2.0 bimanual setup
Test conditions: Different model size and observation setup from the 216.1M LIBERO configuration
Source: TurboVLA project

Isaac GR00T N1

Research paper

Model: Isaac GR00T N1
Version: Public N1 paper model
Hardware: NVIDIA L40
Precision: bf16
Model size: 2.2B total; 1.34B VLM reported in paper
Batch: Paper-specific
Inference latency: 63.9 ms to sample a chunk of 16 actions
Action frequency: Chunk inference metric; not 16 independent control loops
Control frequency note: Low-level motor control remains separate
Robot / benchmark: Generalist humanoid research model
Test conditions: Published N1 paper, L40 GPU, bf16, 16-action chunk; do not apply directly to N1.7 or other GPUs
Source: GR00T N1 paper

WLA-0

Research result

Model: WLA-0
Version: 2B active-parameter prototype
Hardware: NVIDIA RTX 5090
Precision: Paper configuration
Model size: 2B active parameters
Batch: Paper-specific
Inference latency: 40 ms per inference
Action frequency: ~25 model inferences/s from reported latency
Control frequency note: Robot low-level control remains separate
Robot / benchmark: Research benchmark platforms
Test conditions: Author-reported on RTX 5090; world prediction can be disabled during inference according to paper
Source: WLA-0 paper

Model	Version	Hardware	Precision	Model size	Batch	Inference latency	Action frequency	Control frequency note	Robot / benchmark	Test conditions	Source
TurboVLA	216.1M LIBERO configuration	NVIDIA RTX 4090	Paper configuration; precision should be checked in experiment details	216.1M parameters	1	31.2 ms per policy inference	~32 policy predictions/s	Full robot servo loop not claimed to run at 32 Hz	Benchmark + AgileX Piper transfer	Observation-to-action-chunk policy latency on one RTX 4090; excludes full sensing/network/servo pipeline	TurboVLA project
TurboVLA	0.4B RoboTwin bimanual configuration	NVIDIA RTX 4090	Paper configuration	0.4B parameters	1 / paper inference setup	43.4 ms	~23 policy predictions/s from latency	Action chunk = 50 steps; low-level controller frequency is separate	RoboTwin 2.0 bimanual setup	Different model size and observation setup from the 216.1M LIBERO configuration	TurboVLA project
Isaac GR00T N1	Public N1 paper model	NVIDIA L40	bf16	2.2B total; 1.34B VLM reported in paper	Paper-specific	63.9 ms to sample a chunk of 16 actions	Chunk inference metric; not 16 independent control loops	Low-level motor control remains separate	Generalist humanoid research model	Published N1 paper, L40 GPU, bf16, 16-action chunk; do not apply directly to N1.7 or other GPUs	GR00T N1 paper
WLA-0	2B active-parameter prototype	NVIDIA RTX 5090	Paper configuration	2B active parameters	Paper-specific	40 ms per inference	~25 model inferences/s from reported latency	Robot low-level control remains separate	Research benchmark platforms	Author-reported on RTX 5090; world prediction can be disabled during inference according to paper	WLA-0 paper

Methodology

Latency rows are not ranked as hardware-independent model scores. GPU/NPU, precision, model version, batch size, action chunk and measurement boundary stay visible so 40 ms on one accelerator is not treated as equivalent to 40 ms on another.

Included

Published model, robot, benchmark, deployment or failure evidence that can be tied to a retained source.

Not inferred

Unknown autonomy, latency, trial counts, deployment scale and failure causes remain unknown when the source does not disclose them.

Update policy

Versions and historical records remain distinguishable. A new release does not silently overwrite an older model or evaluation condition.