VLA models
8 models
VLAs learn P(action | image, language). One forward pass per action chunk. No future state prediction. The image encoder is always a ViT; the action head varies between tokenization, diffusion, and flow matching.
OpenVLAopen
Kim et al., CoRL 2024 · Stanford / UC Berkeley
VLA7B paramstokenized actions
BackboneLLaMA-2 7B
Vision encoderDINOv2 + SigLIP (fused)
Action headTokenized (256 bins)
Action space7D delta EEF
Training dataOpen X-Embodiment (970K)
LIBERO avg~79.6%
First open-source generalist VLA. Outperforms RT-2-X (55B) by 16.5% with 7× fewer parameters. Fine-tuneable via LoRA on single A100 in 4–8hrs.
OpenVLA-OFT+open
Kim et al., Feb 2025 · Stanford
VLA7B paramscontinuous actions
BackboneLLaMA-2 7B (OFT recipe)
Vision encoderDINOv2 + SigLIP
Action headL1 regression (continuous)
Action space7D delta EEF
Speed25–50× faster inference than base
LIBERO avg79.6%
Optimized Fine-Tuning recipe. Switches to continuous actions (no discretization). Supports multi-image input and bimanual control. Used as VLA baseline in LIBERO-plus robustness study.
π₀ (pi-zero)partial
Black et al., Oct 2024 · Physical Intelligence
VLA3B + 300Mflow matching
BackbonePaLI-Gemma 3B (VLM)
Vision encoderSigLIP ViT-So400M
Action headFlow matching (300M DiT)
Action space7D or 14D (bimanual)
Training dataπ Cross-Embodiment (proprietary)
LIBERO avg~85%
First flow-matching VLA. Excels at dexterous bimanual tasks (laundry folding, table bussing). Weights open on HuggingFace via openpi repo.
π₀.₅ (pi-zero-5)partial
Black et al., Apr 2025 · Physical Intelligence
VLAflow matchingscaled
BackbonePaLI-Gemma (scaled)
Action headFlow matching
Action space7D / 14D
Training dataScaled cross-embodiment
LIBERO avg85.5% (best VLA in robustness study)
Scaling of π₀ with better open-world generalization. Best-performing VLA in the March 2026 WAM robustness study across most perturbation types.
GR00T N1 / N1.5open
NVIDIA, Mar 2025
VLAhumanoiddiffusion head
BackboneEagle-2 VLM
Vision encoderViT (CLIP-based)
Action headDiffusion transformer (DiT)
Action spaceFull humanoid body
Training dataIsaac Lab sim + real robot
FocusHumanoid whole-body control
Foundation model for generalist humanoid robots. Combines high-level VLM reasoning with low-level diffusion action expert. Weights on HuggingFace.
CogACTopen
Li et al., Nov 2024 · Microsoft Research Asia
VLAcomponentizedDiT head
BackboneLLaMA-2 + VLM encoder
Vision encoderDINOv2 ViT-L + SigLIP
Action headDiT (Small / Large variant)
Action space7D continuous
Training dataOpen X-Embodiment subset
LIBERO avg>OpenVLA by 35%
Separates VLM cognition from action generation. VLM produces semantic tokens; DiT action module generates actions conditioned on those tokens. MIT license.
Octoopen
Ghosh et al., RSS 2024 · UC Berkeley
VLAdiffusion headmulti-embodiment
BackboneTransformer (no LLM backbone)
Vision encoderCNN (lightweight)
Action headDiffusion policy
Action space7D + goal image optional
Training dataOpen X-Embodiment
FocusFirst open diffusion VLA
First fully open-source VLA using diffusion policy. Supports flexible goal specification (language or goal image). Standard baseline for ManiSkill3 evaluation.
RDT-1Bopen
Liu et al., Oct 2024
VLA1B paramsDiT backbonebimanual
BackboneDiffusion Transformer (DiT) 1B
Vision encoderSigLIP ViT
Action headDiT (whole backbone denoises)
Action space14D (bimanual)
Training dataOpen X + bimanual datasets
FocusBimanual manipulation
Entire backbone is a DiT — diffusion runs through the whole model, not just the action head. Designed specifically for bimanual (14D) tasks. #1 on HuggingFace robotics 2024.
WAM models
7 models
WAMs learn P(action, future state | image, language). All take language instructions as input — same as VLAs. The difference is the training objective and backbone (video DiT instead of LLM). Output is identical: 7D or 14D action.
Cosmos-Policyopen
Kim et al., Jan 2026 · NVIDIA
WAMjoint diffusion2B
BackboneCosmos-Predict2 video DiT (2B)
Image encoderVAE (latent grid)
Action headJoint diffusion (25 steps)
Language inputYes — T5 text encoder
Chunk size16 steps
LIBERO avg~98.5%
Latency~2000–5000ms
Fine-tunes Cosmos-Predict2 video DiT for robot control. Jointly denoises future video frames + action latents. 25 denoising steps at inference. Weights on HuggingFace.
Fast-WAMopen
Yuan et al., Mar 2026
WAMimplicitWan2.2-5B
BackboneWan2.2-TI2V-5B video DiT
Image encoderVAE (clean latents only)
Action headAction expert DiT (1B) — 1 pass
Language inputYes — T5 precomputed embeds
Chunk size16 steps
LIBERO avg97.6%
Latency~190ms (real-time)
No video generation at inference — single forward pass through Wan2.2-5B. Proves video co-training matters more than test-time video generation. Fastest WAM.
DreamZeroopen
Zhu et al., Feb 2026 · coined "WAM"
WAMzero-shotWan2.1-14B
BackboneWan2.1-14B video DiT
Image encoderVAE
Action headAutoregressive flow matching
Language inputYes
Key featureZero-shot cross-robot transfer
Speed7Hz real-time (distilled)
First paper to coin the term "WAM". 14B Wan2.1 backbone. Zero-shot policy — no robot-specific fine-tuning needed. Flow distillation enables real-time 7Hz control.
LingBot-VAopen
RSS 2026
WAMautoregressiveasync
BackboneMixture-of-Transformers (MoT)
Image encoderVAE (discrete tokens)
Action headInterleaved video+action tokens
Language inputYes
LIBERO avg98.5%
RoboTwin 2.092%
Async generation — predicts next chunk's video while executing current chunk's actions, hiding latency. Server/client eval architecture for LIBERO and LIBERO-plus.
GE-Actpartial
Liao et al., 2025
WAMflow matchingabs EEF pose
BackboneVideo DiT
Image encoderVAE
Action headLightweight flow-matching decoder
Language inputYes
Action spaceAbsolute EEF pose + Euler angles
Chunk size54 steps at 30Hz
Uses absolute end-effector pose (not delta) — unusual for manipulation benchmarks. Strong on LIBERO-plus lighting and noise perturbations in the robustness study.
UWMopen
Xiao et al., RSS 2025
WAMjoint diffusionlarge-scale
Full nameUnified World Models
BackboneCoupled video + action DiT
Image encoderVAE
Language inputYes
Training dataOpen X-Embodiment (large-scale)
FocusPretraining on diverse robot data
Couples video diffusion and action diffusion in a unified framework. Designed for large-scale pretraining across diverse robot datasets before task-specific fine-tuning.
VLA-JEPAopen
Sun et al., 2026
WAMJEPA pretrainingdirect policy
BackboneQwen3-VL-2B + V-JEPA 2 encoder
Image encoderViT (V-JEPA 2 pretrained)
Action headDirect MLP (single forward pass)
Language inputYes
Chunk size8 steps
LIBERO avg~89%
Latency~200–400ms
JEPA-style pretraining shapes representations; direct policy at inference (no MPC, no video generation). Best of both: JEPA robustness + VLA speed. Used in LIBERO-plus + SimplerEnv.
JEPA / latent world models
3 models
Predict future latent representations — not pixels, not video frames. Use ViT (not VAE). No diffusion. Control via MPC at inference (except VLA-JEPA which uses a direct policy). Goal image or reward signal instead of language in V-JEPA 2-AC and LeWorldModel.
V-JEPA 2-ACopen
Bardes et al., Jun 2025 · Meta AI
JEPAMPC1.2Bgoal-image
BackboneViT-H (1.2B, frozen)
Image encoderViT (no VAE — latent only)
ControlMPC (replan every step)
Language inputNo — goal image required
Training data1M hrs internet video + 62hrs robot
Real-robot success65–80% zero-shot
Latency~16s per action (MPC)
Predicts future latent representations, not pixels. Zero-shot to new labs without retraining. Slow due to MPC replanning — not suitable for fast real-time tasks. Weights on HuggingFace.
LeWorldModelopen
LeCun + Mila, Mar 2026
JEPA15M paramsSIGRegMPC
BackboneLightweight ViT (15M total)
Image encoderViT (no VAE)
ControlMPC
Language inputNo — reward signal
Key innovationSIGReg — prevents collapse
Speed48× faster than V-JEPA 2-AC
HardwareSingle GPU, hours to train
LeCun's proof-of-concept that world models don't need scale. SIGReg (Sketched Isotropic Gaussian Regularization) solves representation collapse. Early research stage.
VLA-JEPAopen
Sun et al., 2026 (also in WAM section)
WAMJEPA pretrainingdirect policy
PositionBridge between WAM and JEPA
Language inputYes — standard VLA interface
ControlDirect policy (no MPC)
Latency~200–400ms (VLA speed)
Unlike V-JEPA 2-AC and LeWorldModel, VLA-JEPA takes language instructions and runs as a direct policy — combining JEPA's robustness priors with VLA's speed and language grounding.