Vision-Language-Action · Robotics · Multimodal Models
RT-2 vs π0 vs Qwen-VLA: How VLA Models Output Actions, With Real Numbers
RT-2 wrote actions as text tokens, π0 switched to ~50Hz flow-matching chunks, and 2026 models add 3D reasoning or world-model priors. Simulation is nearly saturated; real-robot margins and OOD shifts differentiate.
The one design question that splits this field
Every vision-language-action model answers the same question differently: how does a language model, which produces discrete tokens, drive continuous motors? The history of VLA models is a sequence of answers to that interface problem, and each answer trades semantic breadth against control precision.
RT-2, from Google DeepMind, chose the simplest possible answer. Discretize gripper translation, rotation, and open/close into bins, write each bin as a text token, and let a web-pretrained vision-language model emit actions the same way it emits words. One model maps image plus instruction to either a sentence or a motor command, and because actions live in the same token space as language, web semantics transfer for free. The cost shows up at the hardware boundary: discretized tokens are coarse for contact-rich motion, and autoregressive decoding is slow on a robot that needs to react.
π0, from Physical Intelligence, chose the opposite answer. Keep a pretrained VLM as the semantic backbone, then bolt on a separate action expert trained with flow matching, the continuous-generation technique behind modern diffusion models. The expert emits action chunks, short bursts of future motor commands at roughly 50Hz, which is the control rate dexterous manipulation actually needs. π0 pays for this with architecture complexity: the action stream is no longer text, so the “web knowledge flows straight into actions” argument becomes “web knowledge conditions a separate controller.”
The 2026 generation refines both answers rather than replacing them. Qwen-VLA keeps a diffusion-style decoder (a DiT action head) but adds embodiment prompts so one checkpoint serves manipulation, navigation, and trajectory prediction. MolmoAct2 keeps action tokenization (its OpenFAST tokenizer) but inserts explicit 3D spatial reasoning between perception and actuation. World Pilot takes a step to the side: it is not a policy at all, but a pair of steering priors injected into a frozen VLA at inference time.
Key numbers
| Model | How actions are emitted | Headline reported result | What the comparison basis actually is |
|---|---|---|---|
| RT-2 | Discretized gripper pose as text tokens | ~6,000 robot eval trials; emergent generalization to novel objects over RT-1, no single success score | vs RT-1 only, same robot data |
| π0 | Flow-matching action expert, ~50Hz action chunks | Laundry folding, table bussing, box assembly on real robots; no benchmark table | Qualitative demos only |
| Qwen-VLA | DiT diffusion action decoder plus embodiment prompts | LIBERO 97.9%, Simpler-WidowX 73.7%, RoboTwin-Easy/Hard 86.1%/87.2%, R2R 69.0% OSR, RxR 59.6% SR | Own protocol, one unified checkpoint |
| MolmoAct2 | OpenFAST tokenizer with explicit 3D reasoning first | LIBERO 98.1% vs π0.5 96.9% (fine-tuned Think variant); DROID 87.1%, +38.7 points over runner-up; +10.6 on MolmoSpaces, +3.2 on MolmoBot | π0.5 on shared simulated evals |
| World Pilot | Not a policy: latent and action steering priors into a frozen VLA | LIBERO-Plus zero-shot OOD 84.7% vs 80.5% base with identical policy weights; steering alone 83.7 (latent) and 83.1 (action) | Same policy weights with and without priors |
Read the rightmost column before the numbers. The strongest rows in this table are the ones where the comparison harness is genuinely shared: MolmoAct2 against π0.5 on the same simulated suites, and World Pilot against its own wrapped policy with identical weights. The headline cross-paper pair people search for, Qwen-VLA 97.9% versus MolmoAct2 98.1%, comes from different papers, different variants (one instruct, one fine-tuned Think), and quite possibly different evaluation splits. Treat it as “both are near the LIBERO ceiling,” not as a one-point victory.
The cleanest controlled result belongs to World Pilot. The same ABot-M0 policy scores 80.5% on LIBERO-Plus without priors and 84.7% with them, so the +4.2 gain is attributable to the world-action priors rather than to a better network. Its own ablation sharpens the story: latent steering alone adds +3.2, action steering alone adds +2.6, and the two combine to +4.2, close to the sum, so the perception-side scene prior and the action-side motion prior fix largely different failures.
The most dramatic number belongs to MolmoAct2. On real-world DROID it reports 87.1% success, +38.7 points over the runner-up. The site papers’ own read of that figure is worth repeating: the margin shrinks to single digits in simulation (+10.6 on MolmoSpaces, +3.2 on MolmoBot, about 1.2 points on LIBERO against π0.5), which is the pattern you would expect if explicit spatial reasoning buys the most exactly where distribution shift is worst, meaning real hardware.
Two absences matter as much as the entries. π0 reports no benchmark table at all; its evidence is breadth of real chores. RT-2 predates the shared benchmark culture and reports trial volume (~6,000 evaluations) plus emergent generalization over RT-1 rather than a success percentage. Neither can be placed on the LIBERO axis, and any ranking that puts them there is inventing data.
How to read the LIBERO-family numbers
Three of the five papers report some LIBERO variant, and each means something different. Qwen-VLA’s 97.9% is standard LIBERO under its own unified-checkpoint protocol. MolmoAct2’s 98.1% is the fine-tuned Think variant, compared against 96.9% for π0.5 in the same fine-tuning setting. World Pilot’s 84.7% is LIBERO-Plus, a zero-shot out-of-distribution benchmark with appearance, camera, and embodiment shifts, run against the same policy weights at 80.5%.
These are not three points on one axis. Standard LIBERO is close to saturated for top models, so differences there mostly measure evaluation noise and protocol choices. LIBERO-Plus scores are lower across the board because the benchmark is deliberately hostile, and World Pilot’s own numbers show the unevenness: 98.6% under lighting shifts but 60.6% under robot-embodiment shift. A model can top the LIBERO-Plus total while one category remains unsolved.
When to use which
- You want semantic generalization on a tabletop arm and can tolerate coarse control: the RT-2 recipe, actions as text tokens, is the simplest to build on top of an off-the-shelf VLM, and its emergent instruction following was demonstrated at ~6,000-trial volume. Avoid it for contact-rich or force-sensitive tasks.
- You need dexterous, high-frequency real-world manipulation: the π0 recipe, flow-matching action chunks at ~50Hz, is the design that folds laundry and buses tables. Budget for a separate action expert and for fine-tuning on new skills; the paper itself notes that truly novel skills still need demonstrations.
- One checkpoint must cover manipulation plus navigation plus trajectories: Qwen-VLA is the only entry here with navigation scores (R2R 69.0% OSR, RxR 59.6% SR) from the same weights as its LIBERO 97.9%. Its instruct variant is competitive rather than dominant on any single suite; breadth is the claim.
- You deploy on real robots and can afford explicit reasoning before acting: MolmoAct2 reports the largest real-world margin (DROID 87.1%, +38.7 over the runner-up) and releases its datasets, which matters for reproduction. Its Molmo2-ER reasoning model averages 63.8% across embodied reasoning suites, reported above GPT-5 and Gemini Robotics ER.
- You already have a VLA policy and want cheap hardening: World Pilot adds no training of the policy at all. A video-pretrained world model that was never action-trained still lifts the wrapped policy to 82.6% on LIBERO-Plus, so most of the gain comes from plain video prediction. The cost is latency: a world-model forward pass per control step.
Limits and open questions
No paper in this set evaluates two of these models under one protocol. The cross-paper numbers above come from different simulators, task suites, fine-tuning regimes, and success-rate definitions, and the table says so per row. A proper VLA comparison benchmark, same robots, same tasks, same metrics, across even two of these systems, does not exist in this literature.
Simulation saturation compresses differences at the top. When standard LIBERO scores cluster in the high 90s, leaderboard deltas stop measuring capability, and the informative signals move to OOD suites like LIBERO-Plus, where World Pilot still drops to 60.6% on embodiment shift, and to real-robot deployment, where MolmoAct2’s DROID margin is the standout and where Qwen-VLA reports 26.6% zero-shot on the dynamic-manipulation DOMINO benchmark. The Qwen-VLA page’s own read is blunt: dynamic manipulation on unseen tasks is mostly unsolved, and unification does not fix it.
Latency and cost are systematically underreported across the field. World Pilot is explicit that its per-step world-model forward pass limits high-frequency reactive control. π0’s flow expert runs at ~50Hz but the paper does not foreground wall-clock cost either. Qwen-VLA does not report parameter count or inference latency in its abstract-level claims, so the practical price of one big multi-task model versus several small specialists is unknown. None of the five answers the question a deployment engineer actually asks: success rate per dollar per hour on my hardware.
FAQ
How do VLA models output actions: tokens, diffusion, or flow matching?
Three interfaces dominate. RT-2 discretizes the action into bins and writes them as text tokens. Qwen-VLA and MolmoAct2 keep token-style or diffusion-style decoders (a DiT head, the OpenFAST tokenizer). π0 generates continuous action chunks with flow matching at ~50Hz, which is the design chosen for dexterity.
Is Qwen-VLA’s 97.9% on LIBERO better than MolmoAct2’s 98.1%?
Not in any measured sense. The numbers come from different papers, different model variants (instruct versus fine-tuned Think), and possibly different evaluation splits. Both sit near the standard-LIBERO ceiling, where deltas mostly reflect protocol choices rather than capability.
Why is World Pilot’s 84.7% on LIBERO-Plus lower than everyone else’s LIBERO score?
Because LIBERO-Plus is a zero-shot out-of-distribution benchmark with appearance, camera, and embodiment shifts, not a clean-condition suite. Its 84.7% is measured against the same policy weights at 80.5%, so the +4.2 is a controlled gain from world-action priors, which is the comparison that matters.
Which VLA model has the best real-robot results?
On the numbers reported in these papers, MolmoAct2’s DROID result (87.1% success, +38.7 points over the runner-up) is the strongest real-robot claim, and its own analysis notes the margin shrinks to single digits in simulation. π0 demonstrates harder dexterous chores qualitatively without a success rate, so the two are not directly comparable.
What is the most reproducible VLA model in this comparison?
MolmoAct2 releases three datasets totaling 3.3M samples, including 34.5k bimanual demonstrations and DROID/SO-100 collections, which makes its fine-tuned results the most checkable. π0’s cross-embodiment pretraining is heavy enough that the paper itself implies reproduction is out of reach for most labs.
Do any of these methods fix dynamic or contact-rich manipulation?
Not according to their own numbers. Qwen-VLA scores 26.6% zero-shot on DOMINO dynamic manipulation, and World Pilot drops to 60.6% on the robot-embodiment axis of LIBERO-Plus. RT-2’s text tokens were never designed for force feedback, and π0’s dexterity demos, while impressive, come without reliability statistics.