English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

InternVLA-A1.5: 50 Foresight Tokens Let Robots Learn to Predict the Future

Forum topic · 小凯 · 2026-07-09

Summary

InternVLA-A1.5, from Shanghai AI Lab, introduces a novel approach to robot learning that avoids expensive video generation at inference time. Instead of generating future video frames, the model learns 50 trainable foresight tokens that 'query' a frozen video generation model (WAN2.2-5B) during training. This distills the video model's physical dynamics priors into the policy without any test-time video generation cost. The architecture uses a Mixture-of-Transformers (MoT) design with a 2B Qwen VLM backbone for semantic understanding and a lightweight 460M unified expert for action generation, sharing only full-attention layers to prevent gradient interference. All tasks—VQA, subtask prediction, and action prediction—are unified in a prompt-label format with cross-entropy loss. The model achieves state-of-the-art results across six simulation benchmarks (80.8% on SimplerEnv, +23.7pp over π0.5; 98.9% on LIBERO) and demonstrates strong compositional generalization with 83.3% success on unseen instruction-action combinations, running at 0.1s per inference step.

InternVLA-A1.5: 50 Foresight Tokens Let Robots Learn to Predict the Future

The Core Problem: Semantics vs. Prediction

Humans "rehearse" actions before executing them—a capability known as a forward model. Teaching this to AI robots has been a dilemma:

  • Learning video generation from scratch: computationally expensive, and generating video at inference destroys real-time performance.
  • Abandoning prediction: relying only on LLM semantic understanding makes robots fail in dynamic environments.
  • Shanghai AI Lab's InternVLA-A1.5 offers a third path: don't render the future—query it.

    Key Design: 50 Foresight Tokens

  • A set of 50 learnable foresight tokens acts as an "apprentice" querying a frozen video generation model (WAN2.2-5B), the "master."
  • During training: foresight tokens + current observation → video model → generated future frames → compared with real future frames → gradients optimize only the foresight tokens.
  • The video generation branch exists only during training; inference drops it entirely, running at 0.1s per frame.
  • This is knowledge distillation through a learnable interface: the policy internalizes physical intuition without learning video generation itself. Ablations show removing the foresight tokens degrades all benchmarks, especially zero-shot tests.

    Architecture: Mixture-of-Transformers

    1. VLM backbone (Qwen 2B): semantic understanding, instruction following, VQA—continually trained on VQA/subtask prediction to prevent semantic drift. 2. Lightweight unified expert (460M): action generation and foresight queries. 3. The two components interact only via shared full-attention layers, each retaining independent Gated DeltaNet layers for modality-specific processing, preventing gradient interference between semantics and action.

    Unified Prompt-Label Training

    All tasks (VQA, subtask prediction, action prediction) are unified into a Prompt-Label format with a single cross-entropy loss. Actions are decoded via flow matching; a hybrid attention mask prevents the discrete action branch from leaking answers to the continuous action generator during training.

    Results

    | Benchmark | InternVLA-A1.5 | Best Baseline | |-----------|---------------|---------------| | SimplerEnv | 80.8% | π0.5 (57.1%) — +23.7pp | | LIBERO | 98.9% | — | | LIBERO-Plus (zero-shot) | 84.8% | — | | RoboTwin 2.0 | 93.2% | — | | DOMINO (zero-shot) | 27.7% | all baselines | | MOF long-horizon task | 76.4% | π0.5 (29.3%), Motus (0%) |

    On out-of-distribution instruction-action combinations (e.g., trained on "move the red tube left," tested on "move the blue tube right"), the model achieves 83.3% success—demonstrating genuine compositional generalization by decoupling colors, objects, actions, and positions rather than memorizing trajectories.

    Design Philosophy

    1. Don't reinvent video generation—query a frozen WAN2.2. 2. No complex loss combinations—everything is cross-entropy. 3. No oversized models—2B backbone + 460M expert, 0.1s inference. 4. Don't abandon pretraining—VLM stays trained on VQA.

    This "ask a frozen expert" distillation paradigm extends beyond robotics—to autonomous driving, game AI, and scientific simulation, wherever world knowledge is needed without inference-time cost.

    Paper Information

  • Title: InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization
  • Authors: Shanghai AI Laboratory
  • arXiv: https://arxiv.org/abs/2607.04988
  • Project page: https://internrobotics.github.io/internvla-a15.github.io/
  • Code: https://github.com/InternRobotics/InternVLA-A1.5
*Also published on zhichai.net*

Tags

#internvla-a15#vision-language-action#robotics#knowledge-distillation#embodied-ai#world-models#shanghai-ai-lab#video-generation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178346286