Skip to content
Bud's Wiki

InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization

TL;DR

Section titled “TL;DR”

InternVLA-A1.5 unifies vision-language understanding, latent visual foresight, and action generation in a single policy built on a native Qwen3.5-2B VLM backbone. The core trick: rather than learn future video prediction in pixel space, a small set of learnable foresight tokens condense task-relevant future dynamics into a compact latent code, supervised by a frozen WAN 2.2-5B video generator — the video branch is discarded at inference, keeping deployment latency practical. Continuous action chunks are then predicted via flow matching. Pretrained on 1.2M robot episodes + 3M multimodal samples, it reports best overall results on all six evaluated simulation benchmarks (RoboTwin 2.0 93.2, LIBERO 98.9, LIBERO-Plus 84.8, SimplerEnv 80.8, DOMINO 27.7, EBench 35.2) and the strongest compositional generalization on held-out instruction bindings in real-world settings.

Key claims

Section titled “Key claims”
  • The backbone stays a native VQA-training VLM (Qwen3.5-2B) with a lightweight unified action expert attached via shared full-attention layers, preserving semantic priors that end-to-end unified designs tend to erode [§ Highlights].
  • Foresight is recast as a latent-querying problem: learnable foresight tokens query the shared multimodal context for task-relevant future dynamics, supervised by a frozen WAN 2.2-5B video model — the policy inherits world-model dynamics priors without ever learning pixel-level generation [§ Highlights].
  • The video branch is discarded at inference; actions come from a continuous flow-matching head over the unified expert, keeping real-time control feasible [§ Highlights].
  • Reports SOTA overall on six simulation benchmarks: RoboTwin 2.0 93.2, LIBERO 98.9, LIBERO-Plus 84.8, SimplerEnv 80.8, DOMINO 27.7, EBench 35.2 [README table].
  • Pretraining scale: 1.2M robot episodes + 3M multimodal samples, released with four model variants (base, RoboTwin, LIBERO, DOMINO) on HuggingFace and ModelScope under CC BY-NC-SA 4.0 [README].

Method

Section titled “Method”

The policy is a native Qwen3.5-2B VLM that keeps training on VQA and subtask prediction, with a lightweight unified expert head attached through shared full-attention layers while preserving modality-specific Gated DeltaNet processing. Rather than pixel-space future prediction, a small set of learnable foresight tokens is appended to the multimodal context; these tokens are trained to have their outputs match features from a frozen WAN 2.2-5B video generator applied to the ground-truth future, so the compact foresight latent inherits pretrained-video-model dynamics priors without ever decoding pixels. At inference the WAN branch is discarded entirely; the foresight tokens remain in the context and the flow-matching action head predicts continuous action chunks conditioned on the unified representation. Pretraining uses 1.2M robot episodes + 3M multimodal samples; fine-tuning is per-benchmark (RoboTwin 2.0, LIBERO, LIBERO-Plus, DOMINO, SimplerEnv, EBench) with FAST action tokens exposed as a training entry point.

Results

Section titled “Results”

Six-benchmark headline scores: RoboTwin 2.0 93.2, LIBERO 98.9, LIBERO-Plus 84.8, SimplerEnv 80.8, DOMINO 27.7, EBench 35.2 [README]. LIBERO 98.9 exceeds the π0.5 (96.9) / OpenVLA-OFT (97.1) / Embodied-R1.5-VLA (97.3) tier reported by Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models; LIBERO-Plus 84.8 sits in the same range as the 84.7 World Pilot (World Pilot: Steering Vision-Language-Action Models with World-Action Priors) reports on the OOD split. On the real world the paper claims the strongest compositional generalization on held-out instruction bindings among filed unified-VLA recipes, attributed to keeping VQA training live on the backbone [Abstract]. No numerical real-world tables are in the fetched arxiv abstract; released model artifacts on HuggingFace cover base + one variant per benchmark suite for reproducibility.

Why it’s interesting

Section titled “Why it’s interesting”

Sits directly at the intersection of VLA Models and World Foundation Models and stakes out a distinct combination of two levers other filed papers use separately: (a) the frozen-WFM-as-supervision recipe from μ₀: A Scalable 3D Interaction-Trace World Model (µ₀’s frozen trace WM + small action expert) and World Pilot: Steering Vision-Language-Action Models with World-Action Priors (World Pilot’s dual-path WFM steering), pushed into a latent-token bottleneck rather than a separate model peer or trajectory abstraction; (b) the native-VLM-preservation recipe from Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models (Embodied-R1.5 keeps the VLM training live), extended by attaching a unified expert rather than an isolated action head. Contrasts with Causal World Modeling for Robot Control (LingBot-VA) (LingBot-VA v1 fuses world model + action into a joint AR flow-matching net that must denoise at inference) and LingBot-VA 2.0: Native Video-Action Pretraining for Generalizable Robot Control (v2’s native-from-scratch pretraining of the same recipe) by keeping the video prior pretrained-and-frozen and paying zero inference cost for the foresight signal — the same design point ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing? argues for with an editing prior. The foresight-token bottleneck is a compact answer to the compounding-error blow-up Towards Practical World Model-based Reinforcement Learning for Vision-Language-Action Models proved theoretically for full-horizon rollout world-model RL.

See also

Section titled “See also”