DeepCybo's PhysBrain 1.5 Leads Open Physical AI as TUM, NUS and CosmosMind Push Spatial and Recursive Intelligence

This evaluation covered visual spatial perception, spatial and multi-view understanding, embodied cognition and planning, spatial referring and affordance, and visual trajectory reasoning. PhysBrain 1.5-8B took open-source first in 14 of the 28 benchmarks and second in 10, QbitAI reported. It scored 73.9 on RoboSpatial-Home, 84.0 on Part-Affordance, and 89.8 on RoboRefit. The 2B version scored 66.6 on the same set; it was listed as a reference and excluded from the open ranking, but its score exceeded other non-PhysBrain open models including Hy-Embodied-VLM-1.0 at 66.0, Embodied-R1.5 at 64.9, and RynnBrain 1.1 at 63.1.

DeepCybo organizes the model around what it calls the Physical Loop: observing the world, understanding space and task, judging action consequences, executing, and correcting based on new observations. The company trained it on Ego360 panoramic human interaction data that records surrounding environment, full-body pose, hand motion, and task-level speech. DeepCybo open-sourced the technical report, 2B and 8B model weights, and evaluation toolkit. On Sept. 10, China's National Data Administration held a symposium on embodied intelligence, and DeepCybo attended and presented its human panoramic data and In Context Data Collection paradigm, according to QbitAI.

At ECCV 2026, Angela Dai, a professor at the Technical University of Munich, presented a different route to spatial AI, Leiphone reported. Real 3D scans contain noise and occlusions, and 2D denoising diffusion does not transfer directly because missing geometry from physical occlusion is neither random nor normally distributed. Dai's approach turns physical occlusion into a structural prior instead of trying to eliminate incomplete data.

The work rests on three components. SG-NN, described as inverse self-supervision, encodes sensor views into known empty space, observed surfaces, and unknown occluded regions, then removes observation frames to train on deliberately incomplete inputs while a mask loss ignores unknown space. Seen2Scene uses visibility-guided flow matching and 3D Gaussian splatting with WorldMesh to address mean blur in large blind spots; Leiphone reported it can render environments spanning seven to eight rooms. GenRecon feeds structural priors learned from synthetic scenes back into real reconstruction and points toward active perception for robots, allowing an agent to predict unobserved space and plan the next best viewpoint. In the Q&A, Dai acknowledged that diffusion sampling remains slow and that accuracy-speed tradeoffs persist for industrial deployment.

At the same conference, Shou Zheng, head of NUS Show Lab, presented Show-o and Show-2, which combine autoregressive text modeling and discrete diffusion for visual generation in a single Transformer, according to Leiphone. The architecture avoids both the slow token-by-token image generation of pure autoregressive models such as Chameleon and the difficulty of connecting continuous diffusion models to discrete VLM tokens. Show-o uses mask-and-predict discrete diffusion to recover visual tokens in about a dozen steps.

Show-2 addresses three bottlenecks: mutual reinforcement between understanding and generation through cycle consistency, continuous video generation with a 3D tokenizer and spatiotemporal bidirectional attention, and a dual-path architecture that shares a 3D encoder at the input but separates understanding and generation inside the model, adding a semantic layer for distillation and alignment. The team also developed Q1, a real-time backbone for low-latency inference, and combined cloud models with edge-fine-tuned models. In robotics, it built tabletop dual-arm systems with grippers and dexterous hands, a wheeled mobile base with an active arm, and teleoperation for state-action trajectories. VLA policies predict actions, world models predict pixel-level future changes, and world action models combine both. Cloud frontier models take about 8 to 20 seconds per decision, while locally tuned models offer near-real-time control, according to the report.

Separately, CosmosMind, working with Stanford, Berkeley, MIT, Tsinghua, Peking University, and more than ten other universities and institutions, released MetaRSI-v1, described as the first meta-recursive architecture unifying Model-RSI, Data-RSI, and Harness-RSI, QbitAI reported. It introduces a Loop Kernel that abstracts self-improvement as a closed loop: consume feedback, propose changes across data, harness, and model, submit them to a verifier, and feed results into the next round. Data-RSI extracts learning signals from trajectories; Harness-RSI edits system prompts, skills, MCP, tools, and memory; Model-RSI updates parameters and structure.

In experiments, MetaRSI-v1 improved Qwen3.5-35B-A3B, a 35B total and 3B active model, by an average of 10.9 points on Terminal-Bench 2.1, SWE-bench Pro, a GPQA-Diamond subset, and AIME, with the SWE-bench Pro solve rate nearly doubling. It also improved six frontier models including GPT-5.6 and Claude Opus 5 by an average of 7.3 points on Terminal-Bench 2.1, according to QbitAI. The team listed five laws: verification determines the self-improvement frontier; self-knowledge expires as capabilities change; capability is independent of carrier while cost is not; credibility must be measured by what cannot be written; and loops do not create capability from nothing. CosmosMind also open-sourced RSI-Harness.

CosmosMind said it is exploring closed loops that touch the physical world through programmable instruments, automated laboratories, simulation, bench experiments, and real instruments. It expects the first physical evolution loop may appear in automated laboratories rather than open environments, according to QbitAI.

Editor's Summary

PhysBrain 1.5's release placed an open-source Chinese physical AI model near leading closed models on 28 embodied and spatial benchmarks, while ECCV 2026 presentations from TUM and NUS Show Lab proposed occlusion-as-prior 3D reconstruction and a unified autoregressive-discrete diffusion architecture. CosmosMind's MetaRSI-v1 extended self-improvement to the methods that improve models, reporting gains for both small and frontier systems. The advances point to more capable spatial, multimodal, and physical agents, with latency and compute tradeoffs still limiting real-time deployment.