01 · Inference resources
Resources needed for inference
GPU memory
Weights (parameters × bytes per precision) plus activations, KV cache, and image tokens. Reported values equal the bf16 weight estimate or exceed it by up to about 4 GB.
Latency
VLM forward pass plus action generation. Autoregressive tokens scale with token count (π0-FAST about 750 ms per chunk); flow and diffusion heads scale with denoising steps (π0 10 steps, 73 ms).
Control frequency
Outputting an action chunk at once and executing it in pieces allows control faster than the inference cycle. π0 reaches up to 50 Hz with 73 ms inference, GR00T N1 120 Hz, and Helix S1 200 Hz.
Where it runs
Onboard the robot (Jetson, embedded GPU), a PC next to the robot (RTX 4090), a remote server (A100, H100), or the cloud (TPU, Gemini API). Large VLMs go on servers; fast controllers stay on the robot.
Source: values reported in each model's paper, README, or model card (the card's ‘Inference memory’). The line is a calculated value and excludes activations and KV cache.
Reported inference speed by device class
02 · Memory calculator
How much GPU memory does this model need?
Pick a model or enter a parameter count to compute weight memory per precision, shown alongside reported measurements and typical device memory.
03 · My device
Models that run at a usable speed on this device
Choose a device, precision, and required inference rate to list models expected to fit in memory and reach that rate. Token-based VLAs must emit several tokens per action, so inference rate is ‘token rate ÷ tokens per action’; for flow and diffusion models one inference is one backbone pass plus several denoising steps.
Device specs and sources
04 · Backbone LLM
Backbone LLMs of VLAs
Counts only models with a confirmed backbone LLM family. ‘config’ marks cases where the paper names only the VLM, so the LLM is taken from that VLM's public configuration.
Models and backbones by family
05 · Reinforcement learning
Where reinforcement learning is used
‘Used’ means the model itself is trained or improved with RL; ‘Partial’ means RL is used only in part of the system, such as low-level controllers, RL-generated data, rewards, or value models. Evidence is quoted from papers and official blogs; models that do not disclose their training process are marked Not disclosed.
VLA RL post-training
Further improves an imitation-learned VLA or policy with RL. Manipulation: π*0.6 (RECAP), GR-RL, SyVLA, GEN-1; navigation: VLN-R1, ABot-N0, LightNav-0 (GRPO family) and VLingNav, TIC-VLA, RING, COMPASS (PPO family); quadruped VLA: MoRE (offline RL).
RL controller training
Humanoid whole-body controllers and quadruped loco-manipulation controllers (VBC, RoboDuet, SLIM) are trained in simulation with PPO or similar and distilled into student policies. 1–4 days on one GPU is common; SONIC used 128 GPUs.
RL low-level controller
The VLA stays imitation-learned and delegates locomotion and whole-body execution to an RL controller. Examples: GR00T N1.5, N1.6 (Decoupled WBC), Helix, Helix 02, Ψ0, WholeBodyVLA, and legged-robot VLAs QUAR-VLA and NaVILA.
Data, rewards, values
Trains on data generated by RL policies (π0.7, Rho-alpha, AC-DiT, CrossFormer), or adds reward models, DPO, or value prediction to training (DYNA-1, NORA-1.5, Cosmos Policy).
06 · Mobile base
How the mobile base is moved
Wheeled mobile manipulation
π0, π0.5, Mobile ALOHA, BRS, AC-DiT, and PanoVLA output base velocity (linear and angular, or x, y, yaw) directly from one policy together with arm actions. MoManipVLA is the exception: it keeps a fixed-base VLA as is and sets the base position by optimization.
Navigation VA and VLA
Most output a few relative waypoints that a PD, MPC, or pure pursuit controller converts into velocity commands (ViNT, NoMaD, NavFoM, InternVLA-N1, TrackVLA, ABot-N0). VLM-based models split into those that output discrete actions as text, such as ‘move forward 25 cm, turn 15°/30°, stop’ following the VLN-CE benchmark (NaVid, Uni-NaVid, StreamVLN, FutureNav), and those that output waypoints (NavFoM, InternVLA-N1, ABot-N0, LightNav-0).
Quadruped
In quadruped mobile manipulation, the VLA or high-level policy outputs only body velocity and pose commands, and a locomotion controller trained with RL in simulation handles leg joints (Walk These Ways in QUAR-VLA and MoRE, VBC, SLIM, UMI on Legs). Most quadruped navigation models output only waypoints or velocity commands that the robot's built-in locomotion controller follows; NaVILA trains and attaches its own RL locomotion policy. CrossFormer (Go1) is the only case where an imitation-learning policy outputs leg joints directly.
Humanoid legged
The standard is a hierarchy in which the VLA outputs locomotion commands (vx, vy, yaw) or latent actions and an RL-trained whole-body controller moves the legs (GR00T N1.6, N1.7, WholeBodyVLA, Ψ0, OpenHLM, ω-0). Atlas LBM outputs foot poses executed by MPC; Gemini Robotics 2 does not disclose a separate locomotion controller.
Each cell shows the number of models that move that base type with that method; darker means more. A model moving several base types is counted in each cell. ‘Other’ covers aerial robots and bases moved only in simulation. Classification follows the action-space description in each paper or official document, summarized in one line under ‘Base control’ on the card.
07 · Compare two models
Two models side by side
Pick models with the ‘Compare A’ and ‘Compare B’ buttons on cards or the lists below to place their diagrams and specs side by side. Rows with identical values are dimmed.
08 · Gallery
Model gallery
Filter by category, release date, size, backbone LLM, action head, inference info, paper availability, and base type. Click the paper title at the top of a card to open the arXiv original; see task, data, and training settings under ‘Details’.
No models match. Try resetting the filters.