RESEARCHWatch a child at play. A room becomes a field of possibilities: something to reach, move, turn, or try again. Every action meets a response from the world.
What kind of intelligence can grow from those encounters? This is where Soma’s research begins.
A story about learning from the ground up ↓A 45-second time-lapse of play. The inspiration for our question.
Experience before instruction.
A cup can be named, described, and measured. It can also be grasped, tilted, balanced, or dropped. These possibilities emerge from the relationship between a body and its surroundings. [6]
For robotics, this suggests a starting point: learn across a wide range of actions and consequences, before narrowing the goal. Could that experience become a foundation for tasks we have yet to specify?
Language models learn from a wide record of human expression before adapting to particular uses. We ask what that bottom-up principle could mean for physical action: learn from varied interactions, then fine-tune toward useful skills.
A foundation is shaped by the experience that goes into it. What should a robot encounter next?
Knowing a world includes learning what is possible within it.
Experience connects action to consequence. John Dewey located the value of experience in the connection between what we do and what happens as a result. Activity becomes instructive when its consequences inform what follows. For Curio, this inspires a practical question: how can each recorded interaction make the next choice more informed? [5]
A body gives the world possibilities. Embodied approaches to cognition examine how bodies and their relations with the environment shape thought and perception. Our engineering analogy is concrete: what a robot can reach, the contact it can make, and the consequences of that contact all matter to the experience it learns from. [6]
A goal needs a way into action. The philosophical discussion of knowing-how, shaped by Gilbert Ryle, asks how practical ability relates to knowledge of facts. For Soma, the useful question is how a language instruction can call on learned physical capabilities. [7]
The child’s play has interests and purposes of its own. Our inspiration is the richness of those encounters. Whether similarly broad experience transfers to later robot tasks is an experimental question. These philosophical traditions motivate our research; they do not establish its results.
Language models offer a powerful computational precedent. A model can learn structure from a large body of text before being adapted to particular tasks. That shared foundation can support many different uses. [1]
Pretraining can also prepare a model to respond to tasks described in language or illustrated with examples in its context. The final request draws on a much longer history of learning. [2]
That is the bottom-up principle we want to investigate in robotics: let reusable capabilities grow from broad experience, then connect them to specific goals.
The shared idea is reusable preparation. For language, broad text supports a pretrained model that can be prompted or adapted for many tasks. For robotics, actions and their consequences form the experience; language grounding and post-training connect that foundation to useful action.
Fine-tuning refines and adapts a pretrained foundation. The diagram is a conceptual analogy. It does not imply that every capability first appears during fine-tuning, that Soma has collected a complete physical distribution, or that these target skills have been demonstrated. Coverage and transfer remain to be measured.
But a robot’s experience has a different shape. Contact can slip. A drawer can resist. The same movement can produce different outcomes as the scene changes. A foundation for action needs experience of those consequences.
And every physical interaction has a cost. Someone—or something—must decide what happens next.
This turns pretraining into a second problem: how do we acquire the experience worth learning from?
Curiosity gives exploration a direction: what might the next interaction reveal? Curio uses an agent to help select those encounters. The recorded experience becomes material for robot pretraining, followed by adaptation to language-directed tasks.
Human operators control robots during real-world collection. Language grounding and task post-training prepare the model for execution.
A fixed task gives collection a destination. Open-ended exploration gives it room to encounter useful transitions along the way: a change of contact, an unfamiliar object state, a failed attempt, a route back.
Learning from play and task-agnostic exploration already have a research history. Soma’s question is how to make this process more deliberate, scalable, and useful for subsequent learning. [3] [4]
Our approach, provisionally called Curio, brings together three parts.
Intrinsic rewards encourage exploration without requiring a separate task reward for every behavior. Simulation lets us vary scenes and initial states, run many interactions, and study what kinds of experience emerge.
The role: a scalable space for exploration and pretraining.
An exploration agent helps human operators decide what to investigate. Memory brings previous attempts into the decision. A world-model component forms hypotheses about what an interaction might reveal. Human curiosity and judgment help choose useful, varied encounters with real objects.
Human operators control the robots during collection. The agent guides and organizes the exploration, retaining observations so the next collection round can build on the last.
The role: real contact, human guidance, and an accumulating history.
Our intended training path starts with simulation pretraining, then uses the real-world collection to ground and extend that foundation. Language grounding and task-specific post-training connect physical behaviors to descriptions and goals, then refine execution for deployment.
The role: a path from broad experience to language-directed action.
Each part answers a different question: what can be explored at scale, what matters in the real world, and how that experience becomes useful. The contribution of each part has to be measured.
The collection agent and the deployed robot policy play different roles. Agent memory helps choose future interactions; recorded interactions form the training corpus.
Detailed training flow ↗Imagine a cup that has only been pushed near its base. Should the next interaction repeat that motion, or investigate what happens closer to the rim? Walk through how an agent could make that choice.
A proposed collection loop, shown through a scripted example. This interface illustrates intended behavior; it is not a recording of a deployed agent or a validated world model.
An action becomes a lesson when its consequences change what happens next. In our proposed collection loop, a recorded failure can prompt a new hypothesis, and that hypothesis can guide another interaction. The test is whether this accumulated history helps robots learn subsequent tasks with less additional data.
An instruction names a goal. A learned foundation must give that goal a path into action. We propose three tests: language steering, composition, and recovery.
Change only the instruction. Test whether the action changes with it.
Combine known interactions into a held-out order with a language goal.
Introduce a controlled disturbance. Observe recovery, retries, and intervention.
Proposed capability tests, not demonstrated results. No new demonstration or test-time weight update is intended for the language-only condition. Prior task exposure must be disclosed.
“Place the red cup inside the tray.” The instruction is short. The preparation behind it is not: seeing the cup, choosing a contact, moving it, and recovering if it slips. Our goal is for language to call on an already learned foundation of physical behavior.
These three proposed demonstrations make that goal concrete: choose a behavior from language, combine familiar skills, and recover when the world changes.
Select an instruction and starting layout within the declared task scope. Record the full attempt, with human intervention and retries visible.
The value of past experience should appear in what becomes possible next. We will test whether exploration improves held-out task learning, how broadly it covers physical interactions, and what the agent contributes.
Reported with scope, trial counts, and uncertainty.
Measured in demonstrations, time, and compute.
Including setup, recovery, and quality control.
The decisive result would be a better learning curve on held-out tasks at a matched total budget. We will isolate simulation, human exploration, agent guidance, memory, and prediction so that an improvement can be attributed to the component that earned it.
| Comparison | Keep fixed | Measure |
|---|---|---|
| Post-training only vs. interaction pretraining + post-training | Downstream data, model, and evaluation; disclose added pretraining cost | Task success, then repeat at matched total budget |
| Fixed-task vs. open-ended collection | Total collection cost | Downstream adaptation |
| Human-only vs. agent-guided | Time, scenes, operator conditions | Useful coverage & transfer |
| With / without memory and prediction | Agent backbone & collection budget | Component contribution |
| No sim / control sim / intrinsic sim | Real data & adaptation budget | Real-world success |
We will distinguish new object instances, environments, language expressions, task compositions, and genuinely new skills. A familiar task initiated by one instruction is not evidence of learning a new skill from one instruction. Data overlap and task-specific post-training will be disclosed.
Human operators control the robots during real-world data collection. The exploration agent assists in selecting and organizing interactions. Agent-guided collection and autonomous robot execution at evaluation are different conditions and will be labeled separately.
In this concept, the component forms hypotheses about object interactions to inform exploration. Its precise implementation, predictive accuracy, and effect on collection remain to be documented. The walkthrough does not establish a validated physical dynamics model.
The independent benefit of simulation, the contribution of memory and prediction to collection, and held-out language-following performance all require controlled experiments. Synthetic chart values and scripted examples on this page are placeholders, not research findings.
The task begins with an instruction.
The learning begins long before it.
The child in the opening film keeps finding another thing to touch, move, and investigate. The scene invites a different first question about intelligence: what has it had the opportunity to experience?
At Soma, we are building toward a growing foundation of physical experience that can support many future goals. Our research asks how to gather that experience, how to learn from it, and how much it changes the next task.
Our ambition is to bring frontier reasoning into contact with the world it describes. Through Curio, we aim to give it a foundation for physical action, learned from interactions and refined for useful tasks.
As frontier models advance, we aim to connect their reasoning to a growing repertoire of physical skills. Soma’s intended advantage is the experience behind those skills: real interactions, failures, and recoveries that become a foundation for useful action.
From understanding the world
to participating in it.
Our long-term goal is intelligence that can pursue varied goals in the physical world, adapt when conditions change, and build on prior experience. Curiosity-driven pretraining is our proposed foundation for that direction.
In the proposed system, a frontier model would interpret goals and coordinate actions. Curio’s learned physical capabilities, refined through post-training, would support robot execution. Observations and outcomes would return to the reasoning process so it can decide what to do next.
The research challenge is to turn this connection into reliable behavior across tasks and environments. We will evaluate language steering, held-out transfer, longer task sequences, and intervention rates under changing conditions.
Research direction. Physical AGI is the long-term ambition; generalization and end-to-end frontier-model integration remain to be validated.
Selected real-robot demonstrations. Current runs are task-specific and primarily overfit; language steering is the next milestone.
Grasp, position, and release paper cups into a multi-level arrangement.
CURRENT FOOTAGE · TASK-SPECIFIC Original clip ↗Two hands, small objects, cups, and a hinged lid across a sequence of interactions.
CURRENT FOOTAGE · TASK-SPECIFIC Original clip ↗Language steering with a fixed model: change the instruction and show the resulting change in behavior, including complete attempts and interventions.
These are existing Soma hardware demonstrations from soma00.com. They show physical manipulation on the current setup. As described by the team, these demonstrations are primarily overfit to their tasks.
They do not establish language steering, held-out generalization, a Curio pretraining advantage, or frontier-model integration. Clip-specific control mode, training exposure, trial counts, interventions, and playback speed remain to be documented before treating them as autonomous policy evaluation.
The next proof point will connect a stated instruction to a complete robot attempt. The accompanying evaluation will report success across repeated trials, task exposure, and any task-specific adaptation.