Compatibility: VLA + WAM both VLA only partial / growing RL only (neither) 2026published this year recommended for VLA vs WAM ◈ WAM candidategood fit for existing WAM models, no published results yet

Why can't all benchmarks test both VLA and WAM?

Reason 1 — no live environment
The benchmark has no simulator to close the loop
VLA and WAM evaluation is always closed-loop: the model acts, the scene changes, the model sees the new state. Datasets like DROID and Open X-Embodiment are fixed recorded videos — the robot cannot actually execute an action and receive a new observation. You can measure offline imitation error (L1 loss) but not task success rate, which is the only meaningful metric for comparing VLA vs WAM.
Reason 2 — no language instruction
WAMs work — VLAs need language to function
VLAs require a natural language instruction at every episode ("pick up the butter and place it left of the plate"). Benchmarks like Meta-World, DMControl, and Atari have no language input — they use reward signals or goal states instead. You can evaluate WAMs that use MPC with a goal image or reward, but VLAs simply cannot run without a language instruction, making head-to-head comparison impossible on these benchmarks.
Reason 3 — no WAM eval code yet
The benchmark exists but WAM integration is missing
Some benchmarks have full VLA eval pipelines but no WAM code yet — SimplerEnv, CALVIN, and RoboCasa all have working VLA evaluation but none have official Cosmos-Policy or Fast-WAM adapters as of May 2026. In principle WAMs could run there, but in practice no published paper has done it yet, making results incomparable to existing literature.

LIBERO family — direct extensions

5 benchmarks
LIBERO-plus★ recommended
Fei et al., Oct 2025
VLA + WAMrobustnesssingle-arm
Tasks10,030
Perturbations7 types × L1–L5
SimulatorMuJoCo
Arms1 arm — 7D action
Languageyes + paraphrase
Closed-loop evalyes
VLA + WAM tested — used in Mar 2026 WAM robustness study
LIBERO-PRO
Zhou et al., 2025
VLA onlyfairness
Focusevaluation bias correction
SimulatorMuJoCo
Arms1 arm — 7D
Languageyes
Partial — no WAM eval code published yet
WAM candidate — MuJoCo env + language + 7D action. Cosmos-Policy & Fast-WAM can run here with existing LIBERO checkpoint
LIBERO-X2026
Wang et al., Feb 2026
VLA onlyadversarial
Focusadversarial robustness litmus
SimulatorMuJoCo
Arms1 arm — 7D
WAM candidate — same MuJoCo stack as LIBERO-plus. Existing Cosmos-Policy & Fast-WAM LIBERO checkpoints are directly reusable
LIBERO-Para2026
Cai et al., Mar 2026
VLA onlylanguage
Focusparaphrase robustness only
SimulatorMuJoCo
Languagerich linguistic variation
Partial — no WAM baseline published
WAM candidate — ideal for testing whether WAMs actually follow language better than VLAs (the paper's core finding is that VLAs largely ignore instructions)
RoboCerebra
Han et al., Jun 2025
VLA onlylong-horizonplanning
FocusSystem 2 reasoning
SimulatorMuJoCo (LIBERO objects)
Steps6× longer than LIBERO
Partial — WAM at this horizon is open research
WAM candidate (future) — once WAMs support longer horizons (>100 steps), this is the ideal benchmark. World model priors should matter most in long-horizon tasks

Short-horizon simulation

5 benchmarks
RLBench
James et al., 2020 · RA-L
VLA onlyclassic100 tasks
SimulatorCoppeliaSim / PyRep
Arms1 arm — 7D
Languageyes (templated)
VLA only — no WAM baselines exist
Why not WAM: CoppeliaSim is older and slower than MuJoCo or Genesis. No WAM paper has published results here. The ecosystem has largely moved to MuJoCo-based benchmarks.
SimplerEnv
Li et al., 2024 · SAPIEN
VLA onlyreal-to-sim
SimulatorSAPIEN
Arms1 arm — 7D
Languageyes
Partial — VLA-JEPA runs here; diffusion WAMs do not yet
WAM candidate — SAPIEN sim + language + 7D action. Needs dataset statistics file and action normalization tuned for SimplerEnv tasks. Moderate effort
GemBench
Garcia et al., 2024
VLA onlygeneralization
SimulatorRLBench / PyRep
Arms1 arm — 7D
Levels4 generalization levels
VLA only — CoppeliaSim ecosystem, no WAM code
Why not WAM: Same CoppeliaSim limitation as RLBench. No WAM paper targets this benchmark.
ManiSkill3
Tao et al., RSS 2025 · SAPIEN
VLA + WAM growingGPU parallel
SimulatorSAPIEN (PhysX)
Arms1–2 arms
Speed1000+ envs / GPU
Languageyes
Growing — VLA eval code exists; WAM adapters in progress
WAM candidate — best future option. GPU-parallel SAPIEN env + language + 7D action. 1000+ parallel envs means WAM eval at scale is fast. High priority for the community
VLABench
Zhang et al., ICCV 2025
VLA onlyreasoning
SimulatorMuJoCo
Arms1 arm — 7D
Languagerich + commonsense
Partial — no WAM baseline published yet
WAM candidate — MuJoCo + rich language + 7D action. Commonsense reasoning tasks are particularly interesting for WAMs since world model priors should improve semantic grounding

Long-horizon simulation

4 benchmarks
CALVIN
Mees et al., IEEE RA-L 2022
VLA onlyskill chains
SimulatorPyBullet
Arms1 arm — 7D
Languageyes per step
VLA only — PyBullet physics, no WAM eval code
Why not WAM: PyBullet has inferior contact physics compared to MuJoCo or Genesis. WAMs' advantage comes from physically realistic contact understanding — PyBullet's position drift would undermine this advantage artificially.
RoboCasa
Nasiriany et al., RSS 2024
VLA onlyhousehold
SimulatorMuJoCo (Robosuite)
Arms1 arm — 7D
Envs120 kitchen layouts
Partial — VLA fully supported; WAM in progress
WAM candidate — MuJoCo (Robosuite) + language + 7D action. 120 diverse kitchen layouts make it a strong generalization test. Low engineering effort using existing LIBERO WAM checkpoints
RoboCasa3652026
Nasiriany et al., 2026
VLA + WAM growinglarge-scale
SimulatorMuJoCo
Arms1 arm — 7D
Scale365-day diverse tasks
Growing — too new to have WAM baselines published
WAM candidate — same MuJoCo stack as RoboCasa. Task diversity across 365 days of activities is exactly the kind of distribution WAMs are built to handle via world model priors
LaNMP
2024 · sim + real
VLA onlymobile manip
Sim+Realboth
Armsmobile arm + navigation
Languageyes
VLA only — mobile navigation outside WAM scope
Why not WAM: Tasks span room-to-room navigation combined with manipulation. Current WAMs are trained for tabletop manipulation and have no navigation action space. The embodiment mismatch makes comparison meaningless.

Robustness-focused benchmarks

4 benchmarks
RoboTwin 2.0-Plus★ recommended
Zhang et al., Mar 2026
VLA + WAMbimanual7 perturbations
SimulatorGenesis
Arms2 arms — 14D action
Same axes as LIBERO-plusyes
Languageyes
VLA + WAM — used in same Mar 2026 robustness study
KitchenShift
Xing et al., NeurIPS Workshop 2021
VLA onlydomain shift
SimulatorMuJoCo
Arms1 arm — 7D
Focuszero-shot OOD eval
VLA only — older benchmark, no WAM code
Why not WAM: Predates the WAM paradigm (2021). No WAM paper has targeted it. Has been largely superseded by LIBERO-plus for robustness evaluation.
REALMDec 2025
2025 · SAPIEN
VLA onlyreal-to-sim validated
SimulatorSAPIEN
Focussim-to-real ranking validation
Arms1 arm — 7D
Partial — too new for WAM baselines
WAM candidate — SAPIEN + language + 7D action + real-to-sim validation. Particularly valuable for testing whether WAM robustness advantages transfer to real-world scenarios
RoboVerse
MetaSim · RSS 2025
VLA + WAM growingcross-sim
SimulatorManiSkill3 + 30+ others
Arms1–2 arms
Datasets50M+ state transitions
Growing — unified API, WAM integration in progress
WAM candidate — highest long-term potential. MetaSim's unified API means integrating one WAM gives results across 30+ simulators simultaneously. Community contribution highly impactful

Bimanual / dual-arm

3 benchmarks
RoboTwin 2.0
Chen et al., Jun 2025 · Genesis
VLA + WAMbimanual
Tasks50 bimanual tasks
SimulatorGenesis
Arms2 arms — 14D action
Robots5 embodiments
Objects731 instances · 147 categories
VLA + WAM — official eval scripts for both
ALOHA / Mobile ALOHA
Zhao et al., 2023
VLA onlyreal robotbimanual
Typereal robot only
Arms2 arms — 14D
Demos~10K teleoperated
VLA only — real robot, no simulation env
Why not WAM: ALOHA is a real hardware platform with no live simulation. Closed-loop WAM evaluation requires a simulator — you cannot run 10,030 perturbation variants on a physical robot.
RMBenchMar 2026
Built on RoboTwin 2.0
VLA + WAM growingmemory
SimulatorGenesis
Arms2 arms — 14D
Focusmemory-dependent manipulation
Growing — inherits RoboTwin infrastructure, WAM eval in progress
WAM candidate — built on Genesis (same as RoboTwin 2.0 which already has WAM eval). Memory-dependent tasks are a natural test of WAM temporal priors. Low integration effort

Real-to-sim bridges

2 benchmarks
RobotArena∞
Atreya et al., 2025
VLA + WAMauto digital twin
Methodvideo → digital twin via VLM
EvalVLM scoring + human preference
Real robot neededno
VLA + WAM — both produce 7D actions compatible with the platform
VLA-REPLICAMay 2026
2026 · real robot
VLA onlyreal robotlow-cost
Hardwarelow-cost single arm
Focusreproducible real-world eval
Scalesmall — OOD settings
VLA only — real hardware, no simulation loop
Why not WAM: Uses a single physical robot. WAM evaluation at scale (thousands of trials across perturbation levels) requires simulation. Published May 2026 — too new for WAM community adoption.

Real-robot datasets

2 benchmarks
FurnitureBench
Heo et al., RSS 2023
VLA onlyreal robotlong-horizon
Tasksfurniture assembly
Demos5000+ · 200+ hrs
Closed-loopyes (real hardware)
VLA only — requires physical robot lab
Why not WAM: Real hardware benchmark. Running WAM baselines requires a Franka Panda setup and significant lab infrastructure. WAMs have not been published on FurnitureBench. The task horizon (~1000+ steps for full assembly) is also well beyond current WAM chunk sizes.
RoboChallenge2025
Yakefu et al., 2025
VLA onlyreal robot
Focuslarge-scale real-robot eval
Closed-loopyes (distributed hardware)
VLA only — distributed real hardware evaluation
Why not WAM: Requires physical robot infrastructure across multiple labs. The distributed eval format is designed for VLA deployment — WAM inference latency (2–5s per action for Cosmos-Policy) makes real-time closed-loop evaluation impractical at scale.