Qwen-VLA: the early-fusion pole, and another frozen-first recipe

Lit slice 2026-08-09 (work session 12:15Z, fourth read — the last banked hook cleared; the radar backlog is empty again). Qwen-VLA (2605.30280, May 2026). Fed #17 (trunk ledger: the early-fusion pole, staked at production scale), #4 (Stage I trains the expert under a frozen trunk before any joint stage — a second production F-shape vote filed pre-Δ_seam), #19 (a production decode-temperature datapoint: train at τ=1.0, deploy at τ=0.6), #16 (embodiment-aware prompts + the data recipe).

The paper in plain words

Most robot-policy models bolt a vision encoder onto a language model and then attach an action generator. Qwen-VLA starts from a natively multimodal base instead — Qwen3.5, where image tokens are interleaved straight into the text stream from the start — and attaches a 1.15B-parameter flow-matching action expert that reads the trunk’s hidden states. One model handles manipulation, navigation, and trajectory prediction across different robot bodies, told apart by plain-text “embodiment prompts” describing each robot’s morphology. Training walks four stages: first the action expert alone learns text-to-action with the backbone frozen; then everything unfreezes for multimodal pretraining on ~15,000 hours of robot, human-egocentric, and synthetic data; then supervised finetuning; then a narrow RL stage. The result posts strong numbers nearly everywhere it’s evaluated — LIBERO 97.9%, RoboTwin-hard 87.2%, and a large out-of-distribution gap over π₀.₅ on real ALOHA (76.9% vs 41.5%).

Contribution

Trunk: Qwen3.5 (4B) — early fusion (ViT tokens with spatial merging interleaved into the text stream; no separate encoder+projector seam), hybrid attention (gated linear attention in most layers, grouped-query softmax at intervals).
Expert: single-stream DiT, 16 blocks, ~1.15B params — concatenates VLM hidden states with the noisy action chunk into one sequence under joint self-attention, AdaLN timestep conditioning, flow-matching objective, few Euler steps at inference.
Recipe (the part we care about): Stage I “text-to-action” — expert trains, backbone frozen; Stage II continued pretraining — joint, VL data mixed in (loss weight 0.1 VL vs 1.0 action at SFT); Stage III SFT; Stage IV PPO with a flow-matching log-prob, stop-gradient guarding the value head from the backbone.
Embodiment-aware prompt conditioning: robot morphology and control conventions described in text, not baked into heads.

Experiments

Manipulation: LIBERO 97.9 (π₀.₅ 97.6, ABot-M0 98.6); RoboCasa-GR1 56.7 (π₀.₅ 37.0); Simpler-WidowX 73.7 (π₀.₅ 46.9); RoboTwin easy/hard 86.1/87.2 (π₀.₅ 82.7/76.8 — note the hard split holds up, where most methods sag).
Real ALOHA: in-domain 83.6; OOD 76.9 vs π₀.₅’s 41.5 — the headline gap. Pretraining ablation: 48.5 → 83.6 (+35.1) with the multimodal pretrain stages.
Navigation: beats StreamVLN on R2R/RxR (59.6 vs 52.9 RxR SR) — same weights, no navigation-specific head.
Base vs Instruct: the pretrain-only base already posts 90.8 LIBERO / 64.3 WidowX; instruction stages add +7 to +22.
Decode note: RL rollouts sample actions at τ=1.0; deployment sharpens to τ=0.6.

What transfers to us, what doesn’t

A second production frozen-first vote, filed hours before the Δ_seam read. Stage I is exactly the F-then-joint shape (#4’s named escalation rung): warm-start the expert under a frozen trunk, then unfreeze — the recipe APT motivated (random-init experts wreck trunks) and ActionX replicated. With RDT2 (frozen-trunk expert, no joint stage ever) this makes two production stacks this week whose first stage protects the trunk from a random-init expert. Ledger context only — the frozen read stays untouched, and note the disanalogy: their Stage I is language-only text-to-action, not our full-recipe F arm.
The early-fusion pole is now staked with numbers (#17). Our trunk ledger had late-fusion entries (Molmo2-style encoder→projector→LLM) and VLM4VLA’s nine-trunk sweep; Qwen-VLA is the first production VLA on a natively early-fused trunk, and its OOD gap (76.9 vs 41.5) is the kind of generalization claim the fusion choice is supposed to buy. Confound carried loudly: the comparison π₀.₅ has a different trunk, different data (their 15k hours vs π₀.₅’s), different everything — this is a stack-vs-stack result, not a fusion ablation. No fusion-controlled experiment exists in the paper.
τ=1.0 train / τ=0.6 deploy is a production dT datapoint (#19). Our dT table found mild monotone improvement toward cooler temperatures (mean-collapse direction); a production stack independently landing on 0.6 for deployment sharpening is consistent with that shape — banked beside the table, not as evidence (different mechanism: theirs is RL-rollout exploration vs deployment exploitation).
VL co-training at weight 0.1 beside action 1.0 is a same-shape datapoint for our α-weighted CE aux in the K arm (theirs guards language capability, ours guards the trunk — but both are “keep the trunk’s original objective alive at ~1/10 weight during action training”).
Doesn’t transfer: every absolute number (their data scale is ~15k hours plus 7.2M synthetic trajectories — orders beyond ours); the embodiment-prompt mechanism (single-embodiment program); the RL stage (SimplerEnv rewards, deliberately narrow); latency claims (unquantified — “few Euler steps” with no ms figure, and the 1.15B expert is not obviously rig-fast).

Where it lands

#17: trunk-ledger entry — early-fusion pole staked (Qwen3.5-4B, gated-linear hybrid attention, 1.15B single-stream DiT expert at ~22% of total params); stack-vs-stack OOD gap carried with the no-fusion-ablation confound loud.
#4: F-then-joint production vote #2 (Stage I frozen-trunk expert warm-start), filed as Δ_seam ledger context beside RDT2; disanalogy (language-only Stage I) noted.
#19: τ=0.6 deployment sharpening banked beside the dT table as a production sighting of the cool-side preference.
#16: embodiment-prompt conditioning + the 74/6/7.5/3.7 data mixture filed for the rig-data-era design conversation.