PhysBrain 1.5: How DeepCybo Unifies Robot Understanding, Action and Prediction

PhysBrain 1.5 is DeepCybo’s open-weight approach to a model that can interpret a scene, generate robot motion and predict what comes next. The key distinction for prospective users is between its reported embodied-understanding results and evidence that a robot can execute tasks reliably in the physical world. 1910

How one architecture handles three tasks

The 2B and 8B versions build on corresponding Qwen3-VL vision-language models. Rather than assigning understanding, action and prediction to separate task-specific heads, PhysBrain 1.5 expresses their outputs as discrete sequences trained through next-token prediction. Those sequences can represent language or spatial answers, end-effector motion, and predicted visual states—including RGB, depth and robot-occupancy information. 1810

That shared format is the point of the design: a scene description, a proposed movement and an estimate of its visual outcome can be learned within one autoregressive architecture. It does not mean a predicted trajectory has been successfully executed by a robot. 18

How DeepCybo trains it

DeepCybo describes two stages. Embodied pretraining draws on task-centred human-interaction videos, pairing scene and task context with reconstructed motion and subsequent observations. Supervised fine-tuning then uses a mixture of human demonstrations, robot trajectories and simulated experience. Calling PhysBrain 1.5 a model trained only on human videos would therefore misdescribe the full pipeline. 112

What the 28-benchmark result shows

DeepCybo reports an average score of 72.5 for PhysBrain 1.5-8B across 28 embodied-understanding benchmarks, ranking it first among the open models in its comparison and first among those models on 14 individual benchmarks. The reported 2B average is 66.6. The suite covers visual-spatial perception, 3D and multi-view understanding, embodied reasoning and planning, grounding and affordances, and visual trajectory reasoning. 191011

In DeepCybo’s published comparison, the 8B score sits near those listed for proprietary GPT-6-Astra (73.3) and Gemini 3.6 Flash (73.0). That is a comparison under the project’s reported evaluation, not an independent ranking of every available model or a measurement of robot-task completion. 118

What teams can verify for themselves

DeepCybo has released 2B and 8B checkpoints alongside its technical report and evaluation resources, giving teams a starting point for inspecting the models and attempting to reproduce the results. 191011

Before relying on the headline average, teams should check per-benchmark results, evaluation settings, possible training–test overlap and whether comparison models received equivalent conditions. For deployment, the more consequential test is closed-loop execution on the intended robot: task success, error recovery, latency and safety with the objects and conditions it will actually encounter. The published understanding score cannot substitute for those measurements. 1910