Simplex full-duplex interaction model: bringing agents into the physical world

Over the past year, we have explored a question: can a flagship language model's reasoning and coding abilities become part of a robot that sees, listens, speaks and acts—and can be interrupted at any moment?

We have built an initial model and tested a proof of concept on a small humanoid robot. This Simplex full-duplex interaction model preview introduces the research direction and what the prototype currently demonstrates.

Bringing reasoning into a continuous scene

Turn-based interaction waits for input to end before generating a response. The physical world does not wait: people interrupt, turn away, and change what a robot needs to do.

We want a model to receive visual and audio information while thinking, speaking and acting. New evidence should be able to enter an ongoing process and redirect or interrupt its output. Full-duplex multimodal interaction means interleaving input and output over time rather than queuing complete turns.

A second goal is limiting alignment tax: adaptation to interaction should preserve the base model's reasoning, coding and common-sense abilities. Initial experiments suggest this may be possible with limited losses; complete before-and-after evaluations will be reported separately.

One training framework, multiple timescales

Our approach combines specially constructed pretraining tasks with a full-duplex time-division multiplexing strategy. We are not detailing the task construction here. The central problem is that robotic capabilities operate at different rates.

We explore a shared pretraining framework for language models, coding agents, embodied navigation and manipulation, social perception, multimodal long-term memory and streaming speech.

Component	Current research role and cadence
Action-command dispatch	Continuous commands at 200 Hz
External audiovisual events	An approximately 200 ms autonomous trigger cadence
Memory and coding	Lower-frequency, task-dependent calls alongside interaction
Training	Exploring end-to-end differentiable optimization through pretraining and post-training RL

200 Hz describes command dispatch, not 200 independent full-model reasoning passes per second. Event-trigger cadence and the time required to respond to an external stimulus are also distinct measurements.

Why we are pursuing this direction

We see VLA and WAM capabilities as components of a broader embodied system. A general robot may also need to read documentation, write code, interpret social situations and remember prior events.

Our research bet is to elicit a flagship model's understanding of the physical world while retaining its digital capabilities, in a form that supports low latency, real-time feedback, scalability and interruption. This is a rationale for our direction, not proof that any model family will inevitably prevail.

Training for combinations that are rarely collected

A single task may combine several people, memory, movement, manipulation and occasional coding. Conventional robot-data collection rarely captures that entire combination.

We have developed an approach to these training requirements and completed an initial model supporting these capabilities. We will share construction details at an appropriate stage. Related work on human-state perception, intent understanding, memory and full-duplex interaction has also produced encouraging results on established benchmarks at relatively low cost; individual evaluations will be reported separately.

A small humanoid proof of concept

In our current prototype tests, a small humanoid combines multi-person interaction, social perception, long-term memory, coding and navigation in one validation system. It dispatches action commands at 200 Hz and has responded to external stimuli within 300 ms in tested scenarios.

The latency is a team-reported PoC observation, not a guarantee across every input, workload or action. Subsequent demonstrations will define timing boundaries, conditions and latency distributions. Whole-body navigation and manipulation across broader tasks still require further validation.

Watching a small, round robot turn toward people, move around and play each day has been rewarding. Bringing an agent's abilities into the physical world—a robot that can act, see, hear, speak and walk—feels like a promising starting point.

Research preview: demonstrations and complete evaluations are in preparation. This describes an initial model and PoC, not a public product release or validation of general-purpose capability.