A 54-author technical report landed on arXiv on September 14: DeepCybo, together with the Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence, released PhysBrain 1.5, an embodied foundation model whose weights (2B and 8B) and evaluation toolkit were open-sourced in the same move on Hugging Face and GitHub. The paper's title states the ambition plainly — "From Vision-Language Models to Physical Foundation Models."

One Model, Three Capabilities

The common pattern in embodied AI today bolts specialized heads onto a VLM: one model stack for understanding, a policy network for action generation, yet another system as the world model. PhysBrain 1.5 takes the opposite route — understanding, action, and prediction share a single autoregressive backbone, jointly optimized under one next-token prediction objective, with no task-specific output heads.

The three capability layers:

  • Embodied understanding: visual-spatial perception, 3D and multi-view reasoning, embodied planning, pointing and affordance grounding, visual-trace reasoning;
  • Action generation: end-effector trajectory chunks predicted via ActionPiece tokens, with a unified action codebook reused across control configurations and robot setups;
  • Future-state prediction: spatially aligned RGB imagery, depth maps, and robot masks that "imagine" the scene after an action executes.

The design maps to the "physical interaction loop" the paper keeps returning to: observations guide reasoning and action, actions change the world, and updated observations feed the next round. Language, action, and visual tokens are encoded as discrete sequences and trained together inside one backbone.

Data and Training: Human Interaction Videos as Pre-training Supervision

The recipe has two stages. Pre-training draws its embodied supervision entirely from human interaction videos — task-centered episodes pairing semantic and spatial context with recovered motion and subsequent observations. The model is then adapted through supervised fine-tuning over a mixture of human demonstrations, real robot trajectories, and simulated experience.

As for the base: PhysBrain 1.5 is built on Qwen3-VL. Starting from a general VLM rather than from scratch also explains another claim the paper stresses — the model retains general multimodal capabilities alongside its embodied skills, rather than trading them away for benchmark points.

Scores: 72.5 Average Across 28 Benchmarks

The technical report evaluates embodied understanding across 28 benchmarks in five categories: foundational visual-spatial perception; spatial and multi-view understanding; embodied cognition, reasoning, and planning; spatial grounding, pointing, and affordance; and visual-trace and trajectory reasoning.

  • PhysBrain 1.5-8B scores 72.5 overall (0–100 scale, unweighted mean across the 28 benchmarks);
  • Among the evaluated open-source models it ranks first on 14 benchmarks and second on 10 (ties included);
  • The paper also draws the comparison: performance on par with leading proprietary models such as GPT-6-Astra and Gemini 3.6 Flash.

One caveat worth stating: the "72.5 average, 14 open-source firsts" figures are self-reported results from the team's technical report. The PhysBrainEvalKit evaluation toolkit is open-sourced alongside, but independent third-party reproduction remains to be seen; the "on par with closed-source frontier" comparison likewise sits inside the official evaluation frame, with no independent cross-vendor benchmark yet. The 2B variant scores 66.6 and is listed for reference only, excluded from ranking.

So What

The embodied-base race right now is about whose token system can unify perception, decision, and prediction. PhysBrain 1.5's answer compresses all three output types into one discrete token vocabulary and uses human videos as free supervision — if this route holds up under independent reproduction, the bolt-on architecture of "VLM plus specialized heads" could fade faster than expected. For researchers, the 2B/8B weights, the evaluation toolkit, and the full technical report are all open, pushing reproduction costs to a low; the next question worth watching is whether the ActionPiece cross-embodiment action codebook stands up on real robots of more morphologies.

References: arXiv:2609.14973 · GitHub: DeepCybo-PhysAI/PhysBrain-1.5