baselines.md

Comparison Methods & Baselines

This document contains baseline methods extracted from experimental tables of major VLA/WAM papers.

VLA Baselines

Method	Paper	Year	Description
DIAL	Decoupling Intent and Action via Latent World Modeling	2026	Intent-action decoupling with latent world modeling
VLANeXt	Recipes for Building Strong VLA Models	2026	VLA architecture recipes
FocusVLA	Focused Visual Utilization for VLA Models	2026	Focused visual attention in VLA
StreamingVLA	Streaming VLA with Action Flow Matching	2026	Streaming with flow matching
ABot-M0	VLA with Action Manifold Learning	2026	Action manifold learning
SimVLA	Simple VLA Baseline for Robotic Manipulation	2026	Simple VLA baseline
Lingbot-VLA	Pragmatic VLA Foundation Model	2026	Pragmatic VLA approach
Gemini Robotics	Bringing AI into the Physical World	2025	Gemini-based VLA
π*0.6	A VLA That Learns From Experience	2025	Experience-learning VLA
X-VLA	Soft-Prompted Cross-Embodiment VLA	2025	Soft-prompted cross-embodiment
UniVLA	Unified Vision-Language-Action Model	2025	Unified VLA with world model
SmolVLA	VLA for Affordable and Efficient Robotics	2025	Compact VLA for edge devices
NORA	Small Open-Sourced Generalist VLA	2025	Small parameter VLA
VLA-0	Building State-of-the-Art VLAs with Zero Modification	2025	Zero-modification VLA
CronusVLA	Efficient Multi-Frame VLA	2025	Multi-frame VLA modeling
OpenVLA-OFT	Fine-Tuning VLAs: Optimizing Speed and Success	2025	Online fine-tuning
AsyncVLA	Asynchronous Flow Matching for VLA	2025	Asynchronous flow matching
AVA-VLA	VLA with Active Visual Attention	2025	Active visual attention
OpenVLA	An Open-Source VLA	2024	First open-source VLA
Octo	Open-Source Generalist Robot Policy	2024	Modular generalist policy
π0	A Vision-Language-Action Flow Model for Robot Control	2024	Flow matching + action expert
AC2-VLA	Action-Context-Aware Adaptive Computation in VLA	Yu et al.	2026
APPLV	Adaptive Planner Parameter Learning from VLA	Lu et al.	2026
Act, Think or Abstain	Complexity-Aware Adaptive Inference for VLA	Izzo et al.	2026
AnyCamVLA	Zero-Shot Camera Adaptation for Viewpoint Robust VLA	Heo et al.	2026
CLARE	Continuous Learning for VLA via Adapter Routing	Römer et al.	2026
DIAL	Decoupling Intent and Action via Latent World Modeling for VLA	Chen et al.	2026
DiT4DiT	Jointly Modeling Video Dynamics and Actions	Ma et al.	2026
EAPruning	Adaptive Pruning with Interleaved Inference for VLA	Huang et al.	2026
ETA-VLA	Efficient Token Adaptation	Wang et al.	2026
FAVLA	Force-Adaptive Fast-Slow VLA	Li et al.	2026
HarvestFlex	Harvesting via VLA Policy Adaptation	Zhao et al.	2026
On-the-Fly VLA	VLA Adaptation via Test-Time RL	Liu et al.	2026
ProbeFlow	Training-Free Adaptive Flow Matching for VLA	Fang et al.	2026
RAFT	Adapting VLA Models via Force-aware Curriculum	Zhang et al.	2026
ROBOGATE	Adaptive Failure Discovery for Safe Robot Policy	Kim et al.	2026
SAMoE-VLA	Scene Adaptive Mixture-of-Experts VLA	You et al.	2026
SCALE	Self-Uncertainty Adaptive Looking for VLA	Choi et al.	2026
SOMA	Memory-Augmented System for VLA Robustness	Li et al.	2026
VGAS	Adaptive Capacity Allocation for VLA	Kim et al.	2026
VLA-Acceleration	Accelerate VLA through Visual Token Caching	Wei et al.	2026
VGAS	Value-Guided Action-Chunk Selection for VLA	Xu et al.	2026
AC-DiT	AC-DiT: Adaptive Coordination Diffusion Transformer	Chen et al.	2025
U-DiT	U-DiT: U-shaped Diffusion Transformers	Wu et al.	2025
VLA-Adapter	VLA-Adapter: Tiny-Scale VLA Paradigm	Wang et al.	2025
A-VL	Adaptive Attention for Large VLA	Zhang et al.	2024
ADEM-VL	Adaptive and Embedded Fusion for VLA	Hao et al.	2024
VL-Adapter	VL-Adapter: Parameter-Efficient Transfer Learning	Sung et al.	2021
RT-2	Vision-Language-Action Models	2023	VLA paradigm establishment
RT-2-55B	Vision-Language-Action Models	2023	55B parameter version
RT-2-1B	Vision-Language-Action Models	2023	1B parameter version
RT-1	Robotics Transformer for Real-World Control	2022	First large-scale robotics transformer
RT-1-35M	Robotics Transformer for Real-World Control	2022	35M parameter version

World Model Baselines

Method	Paper	Year	Description
DreamZero	World Action Models are Zero-shot Policies	2026	World model as zero-shot policy
HCLSM	Hierarchical Causal Latent State Machines	2026	Hierarchical causal world model
WAM	World-Action Model	2026	Action-regularized world model
DreamerV3	Mastering Atari from Pixels	2023	World model RL algorithm
DreamerV2	Mastering Atari from Pixels	2022	Improved world model
DreamerV1	Learning Behaviors from Pixels	2020	Original world model
I-JEPA	Image-based Joint-Embedding Predictive Architecture	2023	Joint embedding predictive
V-JEPA	Video Joint-Embedding Predictive Architecture	2024	Video joint embedding
Cosmos Policy	Fine-Tuning Video Models for Visuomotor Control & Planning	2026	NVIDIA world-action policy

Policy Baselines

Method	Paper	Year	Description
Diffusion Policy	Diffusion for Robot Control	2023	Diffusion model for action generation
ACT	Action Chunking Transformer	2023	CVAE-based action chunking
ACT-ALOHA	Action Chunking with ALOHA	2023	Bimanual manipulation
BeT	Behavior Transformers	2022	Multimodal action discretization
PerAct	Behavior Primitive Discovery	2023	Per-act primitive learning
MVP	Masked Visual Pre-training	2022	Masked visual encoder pretraining
R3M	Universal Visual Representation	2022	Self-supervised visual representation
CQL	Conservative Q-Learning	2020	Conservative offline RL
IQL	Implicit Q-Learning	2021	Implicit offline RL

Dataset Baselines

Dataset	Source	Tasks	Size
OXE	Open X-Embodiment	547	500k+ episodes
RT-1 Dataset	Google Research	9	130k episodes
BridgeData	Berkeley AI Research	8	70k episodes
ALOHA	Stanford	8	10k+ episodes
AgiBot World	Agibot	100+	1M+ episodes
UMI	CMU	6	15k episodes
DROID	CMU	56	80k episodes
Fractal	Berkeley	20	10M+ demonstrations

Benchmark Baselines

Benchmark	Tasks	Evaluation Metric
Libero	Object manipulation	Success rate
Libero-Goal	Goal-conditioned manipulation	Success rate
Libero-Object	Object manipulation	Success rate
RLBench	100+ tasks	Success rate
CALVIN	Sequential manipulation	Success rate
ManiSkill2	Manipulation tasks	Success rate
Metaworld	50 tasks	Success rate

Common Metrics

Metric	Description
Success Rate	Percentage of successfully completed tasks
Average Return	Cumulative reward over episodes
Completion Rate	Tasks completed within time limit
Efficiency	Time taken to complete tasks
Robustness	Performance across variations
Generalization	Performance on unseen tasks/objects
Sim2Real Gap	Performance drop from sim to real
Inference Time	Time per action prediction
Throughput	Actions per second