One-Line Summary

DeepMind unveils a real-time “brain” for robots that coordinates whole-body actions and team tasks, while new research shows how to compress multimodal tokens without retraining and how to test if models truly use visual intermediate states.

LLM & SOTA Models

Gemini Robotics 2 adds whole-body control and teamwork

Google DeepMind introduces a robotics “brain” that lets robots plan multi-step tasks in real time, coordinate with other robots, and control their whole bodies from feet to fingertips. The Gemini Robotics 2 stack brings whole‑body control, advanced dexterity, and multi‑robot collaboration, so a humanoid can walk, crouch, manipulate objects to clean a cluttered room, or team up with other robots to finish faster. ¹

At the center is Gemini Robotics ER 2, an embodied reasoning (ER) model that acts as a high‑level agent: it chats with people, parses a live video feed, plans multi-step tasks, and calls tools or robot interfaces on demand. ER 2 integrates with the Gemini Live application programming interface (API) for bidirectional streaming tuned for low latency, and it’s available to developers through the Gemini API and Google AI Studio, with a private preview in the Gemini Enterprise Agent Platform. ²

New “temporal intelligence” results show ER 2 tracking when to switch steps and when to stop. It reports 57.4% accuracy on progress classification and 91.3% accuracy with a 0.96‑second mean absolute distance on moment finding, delivering this precision at a fraction of the compute cost and 4× the execution speed, with sub‑second latency suitable for physical robots. ²

The stack also adds Gemini Robotics On‑Device 2, a vision‑language‑action (VLA) model optimized to run locally that adapts to new robot bodies in a few hours with typically fewer than 200 examples, plus a new ASIMOV‑Agentic safety benchmark; ER 2 is reported as the safest model to date on safety‑constraint following and human‑proximity tests. ¹

Research Papers

OmniScope: token compression for audio‑video LLMs without retraining

OmniScope speeds up audio‑video question answering by keeping only the most relevant tokens from each modality, without retraining the large language model (LLM). At 25% overall token retention, it achieves up to 3.53× prefill speedup and more than 15% GPU memory reduction, with just a 0.35‑point drop in average accuracy across four audio‑video benchmarks and two Qwen2.5‑Omni model scales. ³

The key idea is to share the user query across modalities but estimate salience separately for audio and video: it allocates modality‑specific token budgets, prunes visual tokens with an anchor‑delta strategy that preserves global context and temporal change, and merges audio tokens within each second to cut redundancy. The authors report best average accuracy across compression settings, and code is available. ³

See2Think: do models really use intermediate visual states?

See2Think is a unified evaluation that checks whether multimodal models truly use sketches, annotations, and intermediate images during reasoning. It introduces See2ThinkBench with 1,200 open‑ended, visually dependent problems across 12 task categories, and Visual Action‑of‑Thought (VAoT), which logs textual thoughts, visual actions, rendered states, and follow‑on reasoning under four controlled inference settings. ⁴

Across representative proprietary and open‑source models, the team finds visual reasoning is strongly model‑ and environment‑dependent with no single setting consistently on top; models often pick relevant visual operations, but faithful rendering is the clearest bottleneck. Under task‑relevant corrupted feedback, accuracy drops by over 10 percentage points, indicating behavioral dependence on visual states and the need for process‑level evaluation, not just final answers. ⁵

Community Pulse

Hacker News (609↑) — Excitement about language-enabled robots mixed with caution about underestimated risks at scale. ⁶

"Honestly having AIs already pretty flawlessly write in basically any language and do good enough translations from any to any is pretty astounding. Throw that on these walking robots and you basically have C3PO. It's not like his arms are good for much." — Hacker News ⁶

"It's really surprising how many people fail to account for scale when determining the risks presented by some of this new tech." — Hacker News ⁶

Why It Matters

Robots that can watch themselves work, time decisions to the real world, and collaborate across different bodies point toward more useful help in homes and workplaces. The reported sub‑second moment timing, progress tracking, and streaming control, combined with on‑device adaptation in a few hours with under 200 examples, lowers latency and integration friction. ²

For builders, two practical levers stand out: control inference cost without retraining by compressing tokens per modality, and measure whether visual tools actually help your model using process‑aware evaluation. Together, OmniScope’s training‑free compression and See2Think’s diagnostics can help teams ship faster, safer multimodal systems. ³⁴