Abstract
Machine Aliveness studies machines that act with initiative, remain grounded in a body, retain history, and live alongside people. This paper examines one part of that agenda: how such machines can continue to develop, and how the process responsible for their development can itself improve. Recursive self-improvement (RSI) requires more than repeated improvement. A change made today must alter the machinery that produces better systems tomorrow. For digital agents, much of that machinery—code, tools, evaluators, data pipelines, and scaffolds—can already be made computational and editable. Physical agents cross a harder boundary: the process that produces their next improvement also depends on a world they do not own.
We call the missing systems layer the Second Loop: the layer at which recursive improvement reaches beyond a learner and into the processes that organize subsequent interaction. The First Loop unfolds in reality and produces committed consequences. The Second Loop can revise the agent, the procedures that construct and evaluate change, and the computational processes through which later worlds are instantiated and allowed to continue. The central question is therefore not merely how to improve a representation of the world, but what kind of computational world can itself become part of recursive improvement.
We propose Agents as World (AAW) as an organization for this world side of the Second Loop. The underlying object is a World Process: a persistent computational process through which participants, events, physical consequences, and histories continue to unfold. In AAW, agents constitute the active organization of that process. Actor Agents participate from local perspectives; World Agents construct, coordinate, and revise ongoing processes; and a tool-grounded World Kernel acts as a commitment layer, separating proposed futures from committed history. For human-centered physical intelligence, people are not only sources of demonstrations, instructions, or feedback. They are persistent participants whose memory, intentions, routines, and relationships become part of the world’s continuing dynamics.
This is one route from machine intelligence toward machine aliveness: not simply more capability, but a persistent developmental process whose own machinery can be revised, inherited, and repeatedly grounded in external reality.
The position
The First Loop unfolds within a world. The Second Loop revises the processes through which later worlds unfold.
1. What makes physical self-improvement recursive?
1.1 From self-improvement to recursive improvement
A system can become more capable without becoming better at producing its next improvement. Weights change, a memory store grows, a prompt is revised, a skill is distilled, or a reward is optimized. These are genuine forms of self-improvement. But they do not by themselves make the improvement process recursive.
We use recursive self-improvement (RSI) in a systems sense: retained outputs from one improvement cycle must alter the conditions, tools, or procedures that determine later cycles. The relevant object is therefore not only a model, but the machinery that produces further capability. A stronger system should change not merely what it can do, but increasingly how its next improvement can be found, tested, and retained.
Reinforcement learning is an optimization method that can operate inside this machinery; self-improvement describes a capability change; RSI concerns whether retained changes reshape the machinery of future improvement. Self-modifying digital agents make the distinction especially visible because code, tools, evaluators, and scaffolds can themselves become revision targets. [1] [2]
For physical intelligence, that machinery necessarily crosses the agent–world boundary. Capability depends not only on the learner and optimizer, but also on the physical and human processes with which the learner interacts, the histories those interactions leave behind, and the mechanisms by which computational worlds are constructed and continued. The world side is therefore part of the systems problem of RSI—even though improving a predictive model of the world is not, by itself, sufficient for recursion.
We write the recursive system at iteration k as:
Here,
1.2 The First Loop is reality
For a physical agent, the First Loop is unavoidable: act in reality, observe consequences, and return with external evidence. Reality gives physical intelligence its purpose, exposes assumptions that internal systems cannot settle on their own, and remains the final place where useful capability must work.
But reality cannot serve as the editable substrate of recursive search. We can change an agent's weights, memory, tools, or execution logic; we cannot fork an entire household from the same instant, restore a person to a prior experience, duplicate a month of life, or rewrite the transition rules of the physical world. Changing a state in reality is not the same as editing the process that generates states.
This is a structural asymmetry rather than a deficiency of reality. Exploration is tied to elapsed time, hardware availability, safety constraints, and human participation. Controlled comparison is expensive because counterfactual conditions cannot be reproduced on demand. The very world that supplies the most authoritative evidence is therefore a poor place to conduct every internal search step.
Simulation-based robotics already separates the scale of training experience from direct physical execution. [3] The RSI question is one level higher: how can the computational machinery that produces learning situations itself become an explicit, revisable, and inheritable part of the improvement system?
1.3 The Second Loop sits above interaction
The Second Loop is that systems layer. It is not simply an agent interacting with a simulator. It operates on the processes that determine how later interaction is organized: the learner, the computational world in which hypotheses are exercised, and the procedures that construct, evaluate, select, and inherit changes can all become revision targets.
The distinction between the loops is functional. The First Loop unfolds through interaction with reality. The Second Loop revises the processes that determine how subsequent interaction can unfold. For physical intelligence, this necessarily includes physical dynamics, human dynamics, temporal organization, and the machinery that instantiates computational worlds without confusing proposed events with committed consequences.
This prevents the thesis from collapsing into “build a better world model.” A predictor can become more accurate while the surrounding improvement process remains fixed. The Second Loop asks for something broader: a computational world whose ongoing processes, participant dynamics, temporal organization, and construction procedures can themselves be inspected, revised, and inherited as part of recursive improvement. The point is not to manufacture more future experience; it is to make the machinery that generates later learning conditions explicit, testable, revisable, and inheritable.
2. The world as a continuing process
2.1 The world side must be executable and revisable
Once the world side enters the systems boundary of physical RSI, the relevant object is no longer adequately described as a passive environment or a predictor of observations. We call the object of interest a World Process: a persistent computational process through which participants, events, physical consequences, and histories continue to unfold over time.
A World Process specifies more than what exists. It specifies which processes continue when nobody is observing them, what each participant can perceive, how physical and human dynamics interact, how time advances, which tools determine executable consequences, and which mechanisms can be revised between world versions. Its implementation may include neural models, programs, scene descriptions, process models, actor specifications, physics settings, state stores, and tool configurations. It need not be end-to-end differentiable. [3] [4]
Revision does not imply that history is negotiable. The Second Loop may change the mechanisms by which later worlds are instantiated, but a process already underway must preserve its committed consequences. A revision creates a new world version; it does not retroactively rewrite what participants already observed. This tension—revisable processes, committed history—will motivate the World Kernel in Section 3.
2.2 The human is inside the loop
Representative generalist embodied systems already use humans extensively—as sources of human video and demonstrations, and as instruction-givers or interactive operators at execution time. [13] [14] These are essential forms of human grounding. In these roles, however, the person largely appears at the boundary of the learning system: supplying data, commands, corrections, or evaluation.
For human-centered physical RSI, that boundary is too narrow. People are part of the world dynamics the robot acts upon, and part of the dynamics that remain present after the interaction ends. A household member does not reset after a task. They remember interactions, form expectations, carry private information, change routines, become more or less receptive, and bring those changes into later encounters.
The human boundary
People are not merely interfaces to the world. They are persistent, adaptive parts of the world.
This changes the modeling role of the human. A demonstrator provides data; an instructor provides a command; an evaluator provides a judgment. A Human Agent inside the world carries state forward and participates in the causal history through which later interaction unfolds. Persistent social agents and shared human–robot worlds have precedents in Generative Agents and Virtual Community. [5] [6] Our concern is how this persistent human state enters the systems problem of physical RSI.
2.3 The world remembers and commits
A continuing world does not reset after each interaction. People remember, objects remain changed, routines shift, processes continue, and unfinished events constrain what can happen next. The world does not merely generate successive states; it carries forward the consequences of its own history.
For human-centered physical intelligence, this is especially consequential. A robot's action can change attention, expectations, routines, preferences, trust, and whether a later interaction is welcomed at all. The learner therefore does not simply act in a world; through interaction, it also changes the conditions of the world it will encounter again.
These consequences at least three coupled forms of dynamics: physical dynamics, human dynamics, and interaction dynamics. Physical distance changes whether a person can perceive an event. Attention changes whether an interruption is noticed. A social response changes the learner's next action; that action can change the physical scene and the future relationship. Treating “physical feedback” and “human feedback” as independent channels loses precisely the coupling that long-horizon physical intelligence must learn to navigate. This is stronger than requiring a richer user profile or a more capable dialogue model. The relevant object is the evolution of a participant in the shared world under interaction. A useful Human Agent must therefore model not only what a person might do now, but how the robot's own behavior can change what that person will later perceive, expect, prefer, and do.
Persistence is therefore part of the RSI problem. A world that participates in recursive improvement must preserve enough of its history for earlier consequences to remain causally available to the future. But a persistent world must also distinguish between what might happen and what has happened. Proposed futures may be revised; realized consequences become part of committed history. The Second Loop may change how later worlds unfold, but it should not silently rewrite the history from which the present world proceeds.
The world must remember what has happened, and commit realized consequences to the history from which what happens next unfolds.
2.4 World-making must expose its own machinery
Finally, the world-making process must expose the machinery by which the world is made. If the recursive system changes only finished scenes or trajectories, it has little leverage over the next generation of learning conditions. Reusable construction procedures, process models, observation interfaces, event rules, human-dynamics models, and temporal policies must themselves be addressable objects.
The world side of the Second Loop must therefore expose not only states and outcomes, but the processes through which later worlds are constructed, continued, and revised.
3. Agents as World (AAW)
Agents as World (AAW) is our proposed organization for the world side of the Second Loop. It is not a synonym for Physical RSI, and it does not claim that agents replace physics, state, or simulation tools. Its claim is organizational: agents can increasingly form the persistent control structure through which a World Process is constructed, operated, inspected, revised, and inherited, while specialized substrates continue to determine executable consequences.
Why agents? The answer is partly ontological and partly technological.
In human-centered physical intelligence, part of the world is already agentic. People are persistent causal participants, not annotations on a physical scene. Their beliefs, memories, intentions, routines, and relationships evolve through interaction and alter how the world subsequently unfolds. A computational world that treats these dynamics as first-class therefore already contains agents as part of what it represents. At the same time, the role of computational agents is expanding beyond representation. Until recently, the machinery required to construct and operate a computational world was largely assembled from outside it: developers authored scenes, wrote event logic, configured simulators, connected policies, and maintained interfaces across specialized systems. Models could generate parts of this stack, but its organization remained mostly external.
That boundary is beginning to move. Frontier agents can increasingly operate executable toolchains rather than merely describe their outputs. They can author environments, invoke simulators and renderers, write and modify programs, inspect resulting state, coordinate other agents, and increasingly act through embodied controllers. GPT-6 Astra, SceneSmith, Agent as Policy, and recent robot-control experiments illustrate different parts of this shift.[15] [16] [17][18]
The important point is not any one demonstration. It is the emergence of a common control abstraction across substrates that were previously organized separately. Physics engines, renderers, databases, event systems, process models, and robot controllers remain specialized execution systems. Agents need not simulate everything themselves. They can instead become part of the machinery that organizes how these systems are composed, invoked, inspected, and revised.
The technological premise
Agents do not need to simulate everything. They need to organize the systems that make a world run.
This is the technological premise behind AAW. AAW asks what becomes possible when agents remain above these heterogeneous substrates as a persistent organizational layer—not only constructing the world, but operating, coordinating, inspecting, and revising it as it unfolds.
Most agentic simulations place agents inside a world. AAW also places agents on the other side of that boundary: as part of the programmable machinery through which the world is made and continued. Agents therefore occupy two distinct roles. Actor Agents participate within the world from local perspectives; World Agents help construct, operate, and revise the processes through which that world unfolds. A shared World Kernel remains responsible for authoritative state and executable consequences.
The resulting lifecycle is deliberately process-oriented: instantiate the world, let it unfold, revise how later worlds are organized, inherit the revision. Recursion enters when retained revisions alter the machinery available to subsequent improvement cycles.
3.1 Actor Agents: local perspectives that persist
An Actor Agent represents a participant whose behavior depends on a local perspective rather than omniscient world state. The abstraction is general: an actor may be a person, worker, visitor, animal, robot, or another decision-making participant. For RSI, the important property is persistence—identity, capabilities, goals, beliefs, attention, memories, and relationships can carry forward so that earlier interactions change later conditions.
Actor Agents must generate more than dialogue. An actor can move, wait, ignore an interruption, collaborate, disagree, demonstrate a behavior, change a plan, or create a new constraint for another participant. Embodiment and observation boundaries remain essential: an actor should not remember an unheard conversation simply because another model generated it, nor complete a physical action merely because it was described in language.
The general Actor-Agent abstraction keeps AAW applicable beyond households. But for the class of physical intelligence we care about most, one subclass deserves explicit treatment: the human.
3.2 Human Agents: persistent human dynamics
A Human Agent is not merely an NPC with a longer prompt. It is a model of a persistent participant whose state can influence—and be influenced by—the physical agent over time. Its specification may include relatively stable characteristics, while its runtime state includes current goals, beliefs, attention, affect, memories, relationships, routines, preferences, and private observations.
The central modeling target is interaction dynamics. It is useful to infer what a person may believe or intend now, but human-centered physical RSI also needs to represent how the robot's own behavior changes later human state. An interruption can alter attention; repeated helpful behavior can establish a routine; an unwanted approach can change later willingness to engage. Those changes persist as part of the shared world history encountered again by the learner.
Human Agents must therefore remain physically situated and epistemically local. What a person can observe depends on position, attention, shared history, and available signals. A Human Agent cannot know an event that never reached them, and a proposed physical behavior must still pass through the same execution substrate as any other action.
They must also remain non-oracular. A simulated preference, correction, demonstration, or emotional response is model-internally generated evidence, not ground truth about a real person's private state. Human Agents are hypotheses to be calibrated against external observations, and making them more compliant to improve the robot's score is not progress in representing human dynamics.
This role is what distinguishes persistent human dynamics from using humans only as data or supervision. The person is part of the causal world state whose evolution helps determine later interaction—and therefore part of the world side that a Second-Loop architecture must preserve and model.
3.3 World Agents: construct, operate, revise
World Agents are responsible for the non-actor world across three stages: construction, operation, and revision. Their purpose in the Second Loop is not merely to generate scenes, but to keep the mechanisms that produce world behavior accessible to later improvement.
During construction, they interpret a world specification, select or create assets, compose layouts, configure physical properties, instantiate processes, and assemble the tools required for execution. A World Agent may use scene generators, physics engines, Blender, code, procedural asset pipelines, or validation utilities; no single simulator backend defines the paradigm.
During operation, their responsibility continues. They maintain semantic process state, schedule events, manage dependencies, react to actor actions, and coordinate tools. A world is not complete when its geometry loads. Appliances run, visitors arrive, tasks are interrupted, queues grow, objects change state, and processes continue outside the learner's current focus.
During revision, agents use new evidence, discrepancies, and changed requirements to modify scene specifications, process models, physical parameters, event rules, or the programs that build and operate later worlds. The target of a revision may be one instance or a reusable method that changes an entire family of later world instances.
This separation is also what makes the architecture scalable. Expensive model reasoning need not occur at every physics step or for every object. Agents can author reusable processes and let ordinary code execute them until a new decision, interruption, or revision is required.
3.4 The World Kernel: revision without rewriting history
Recursive improvement creates a tension: the processes that organize the world must be revisable without making already-observed history negotiable. AAW therefore places agents above a shared World Kernel: the authoritative layer for public state, observation boundaries, event ordering, tool execution, snapshots, branching, and replay.
Agents propose. Tools execute. The world commits. An agent may decide that a person starts walking, a door should open, or a process should begin. The relevant simulator, program, or declared abstraction determines the resulting state transition. Agents can revise the mechanism used by later world versions; they cannot retroactively rewrite a committed consequence inside the current experiment.
A compact implementation view is:
In this view, Gφ is the graph of Actor and World Agents together with their roles, tools, and communication permissions; Bφ is the World Kernel; and Cφ coordinates causal time and activation. The physical agent under improvement is a participant in the resulting world, not an omniscient member of the world-building system.
The Kernel is therefore not only an engineering convenience. It is an epistemic boundary for the Second Loop. Simultaneous actions require arbitration; agents cannot consume state they could not observe; snapshots preserve pending events, active processes, memories, and backend state; reproducible branches pin world mechanisms and randomness. The system can change how future evidence is produced without making present evidence whatever the agents wish it to be.
3.5 Persistent time, selective detail
A recursive physical system must preserve the consequences of its own history long enough for those consequences to matter. AAW therefore separates the continuity of the world from the amount of computation spent describing each moment. The world has one causal timeline with physical meaning; the system may allocate different computational resolutions along that timeline.
A three-second action occupies three seconds of world time regardless of how many model calls or physics substeps are required to compute it. Likewise, a quiet hour need not receive the same dense reasoning as a three-minute human–robot interaction. Represented time and compute time are different quantities.
Coarse process models can advance routine periods, preserving active conditions, memory, and opportunities for interruption. When interaction becomes dense, the relevant interval expands into fine execution: speech, motion, attention, interruption, cooperation, and physical response unfold at the granularity required by the task.
Resolution changes dynamically. An unexpected sound, a robot-initiated approach, or a process failure can immediately open a detailed interval even when the coarse plan predicted no interaction. After fine execution, realized outcomes update future plans.
A strict boundary separates planned futures from committed history. Refinement can resolve what has not happened yet; it cannot rewrite a fact that another participant already observed. The objective is long-horizon continuity without paying fine-grained inference cost everywhere.
3.6 Inheritance: retain how worlds are made
Inheritance is what turns world revision into a recursive capability rather than one-off simulation editing. The Second Loop should retain more than finished scenes and trajectories; it should retain the methods that make future scenes, processes, and experiments possible.
Inherited artifacts can include observation interfaces, interruption models, asset-construction programs, calibrated actor models, process templates, evaluators, temporal policies, or the procedures used to assemble and test world variants. The important property is leverage: a retained improvement changes what later cycles are able to construct or evaluate before those cycles begin.
This creates three distinct levels of change. An actor changing its mind is a state transition. Revising the model that determines how such changes occur is a world-process update. Improving the procedure that discovers, tests, and retains such revisions changes the improvement machinery. RSI concerns the interaction of these levels, not the existence of any one update in isolation.
The learner and the world also have different optimization responsibilities. The learner seeks useful behavior. The world seeks better coverage and a better account of relevant interactions. Making simulated people more compliant is not world-process progress; increasing a difficult scenario's frequency is a curriculum intervention, not evidence that the scenario is common in reality. Social and behavioral validity therefore still requires independent evidence. [10]
The recursive relationship is now explicit: inherited world-making mechanisms change the conditions under which the learner can improve; a changed learner alters what interactions matter; and the higher-order improvement process can revise both before the next cycle begins. Whether this produces sustained capability growth is an empirical question—but the architecture makes that form of recursion expressible.
3.7 A household instantiation: LuvWorld
LuvWorld gives the AAW proposal a household form. The home brings physical presence, persistent agency, and social history into the same setting: people and machines encounter one another repeatedly, and the consequences of one interaction can become the conditions of the next. It is a setting in which physical intelligence cannot be separated from the continuity of its relations.
In LuvWorld, persistent Human Agents carry local observations, memory, goals, attention, relationships, routines, and evolving preferences. World Agents maintain rooms, objects, appliances, activities, schedules, and event processes through simulation and executable tools. The World Kernel keeps physical dynamics and human dynamics on one causal timeline, separating intentions and model proposals from committed state.
Consider an evening in which one person is working, another is preparing to leave, and the robot is deciding whether to initiate contact. The relevant question is not only what response is plausible now. The robot's choice may change attention, interrupt a plan, establish a preference, shift a routine, or alter whether a later interaction is welcomed. Those consequences persist as part of the world history in which later interaction takes place.
This is why the household is more than an application domain for our RSI thesis. It makes the feedback structure of human-centered physical intelligence unavoidable: the learner helps shape the people and social conditions that remain present in its later interactions. LuvWorld is therefore a testbed for recursive improvement in a world whose most consequential state is partly human state. Other domains can instantiate different Actor Agents, World Agents, tools, and process models while preserving the same Second-Loop architecture.
4. Closing the loops: grounding recursive change in reality
Real2Sim2Real as an epistemic bridge
The Second Loop creates an internal space in which the machinery of improvement can change. The First Loop determines whether those changes remain connected to the world the system is supposed to serve. Real2Sim2Real links the two, but it is not itself the definition of RSI. Its role is epistemic: external observations constrain the worlds and agents that are allowed to survive recursive internal search.
Real experience can inform several parts of the system at once. It can update the learner directly—its memory, policy, tools, preferences, or harness. It can also constrain the world side: actor models, environment processes, temporal assumptions, construction programs, and the interfaces through which consequences are measured. The point is not that every observation triggers a new simulator, but that the Second Loop has explicit places where external evidence can revise the processes and procedures used in later internal worlds.
Inside the Second Loop, world variants, learner variants, evaluators, and retained procedures can be compared under controlled assumptions. This creates leverage because many hypotheses about learning can be explored without asking reality to host every branch. But internal scale is useful only if the system preserves the distinction between internally generated evidence and external evidence.
The return path is therefore selective. Success inside an AAW world does not authorize deployment by itself. Candidates still face real constraints, regressions, safety requirements, user expectations, and product-specific release decisions. The Second Loop can accelerate internal search; it cannot promote its own simulations into ground truth.
For household intelligence, the same real interaction may update a user-specific memory, reveal a mismatch in a Human Agent, and expose a missing temporal or social dependency in the current world. These updates have different meanings and should remain distinguishable. AAW makes these different revision targets explicit; the First Loop determines whether the resulting system remains valid for the household it serves.
This is also the bridge between research infrastructure and commercial deployment. Deployed systems reveal which physical, human, and temporal processes matter enough to reproduce computationally, while the Second Loop provides a place to test candidate changes before they become product behavior. Shared tools and world-making procedures can scale across deployments; private household histories remain external evidence governed by consent and appropriate separation.
The resulting position is broader than a simulator, a policy, or a Real2Sim2Real recipe. Physical RSI is the objective. The Second Loop is the systems layer in which the machinery of later improvement becomes revisable. World Process names the continuing computational world. AAW is our proposed organization of that process. Human Agents make persistent human dynamics first-class. The World Kernel separates proposed futures from committed history. Real2Sim2Real is the epistemic bridge that constrains recursive internal change with external evidence.
Machine intelligence asks what a machine can do. Machine aliveness asks what it can become. Our proposal is that physical RSI requires a Second Loop in which the machinery shaping that becoming—agent, world-making process, evaluators, tools, and retained procedures—can itself be revised, while reality remains the final source of external constraint.
Research context
The Second Loop sits at the intersection of several established lines of work. Simulation-based robotics and learned world models separate learning experience from direct physical execution, while open-ended systems such as POET demonstrate that environments and solutions can be generated, selected, and inherited together. [3] [4] [11] These lines establish that the conditions under which learning occurs can themselves become computational objects rather than fixed backdrops.
Persistent social agents point to another part of this picture. Generative Agents and Virtual Community model participants whose memories, activities, and interactions continue over time, while generalist embodied systems increasingly draw on human video, demonstrations, language instructions, and interactive redirection. [5] [6] [13] [14] For human-centered physical intelligence, these threads meet when human state is treated not only as supervision for the learner, but as part of the continuing dynamics of the world the learner repeatedly encounters.
A parallel line of work makes the machinery around the learner increasingly programmable. Tool-using and self-modifying agents show how executable programs, scaffolds, evaluators, and other components of an improvement process can themselves become objects of revision. [1] [2] [8] [9] More recent systems extend this agentic interface across previously separate substrates: software authoring, 3D world construction, simulation, and embodied control. [15] [16] [17] [18] We take this convergence as a technological premise for AAW: specialized simulators, renderers, databases, and controllers remain the substrates of execution, while agents increasingly provide an organizational layer across them.
Recent work on Physical RSI has similarly argued that open-world self-improvement cannot concern the actor alone: unexpected outcomes may expose limitations in either the actor or its computational account of the world, requiring both to remain open to revision. [12] This places the world side of physical intelligence inside the broader problem of recursive improvement.
Taken together, these lines motivate the systems object developed in this paper: a World Process that continues through time, carries participants and consequences forward, and exposes the machinery by which later worlds are constructed and revised. The Second Loop names the layer in which this machinery enters recursive improvement; Agents as World gives that process an active organization; and the World Kernel preserves the distinction between revisable world-making mechanisms and committed causal history.
References
[1] Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
Jenny Zhang et al. · Research paper · 2025.
[2] Hyperagents
Jenny Zhang et al. · Research paper · 2026.
[3] Learning dexterity
OpenAI · Research article · 2018.
[4] World Models
David Ha and Jürgen Schmidhuber · Research article · 2018.
[5] Generative Agents: Interactive Simulacra of Human Behavior
Joon Sung Park et al. · Research paper · 2023.
[6] Virtual Community: An Open World for Humans, Robots, and Society
Qinhong Zhou et al. · Research paper · 2025.
[7] SimPy: Overview
SimPy documentation · Discrete-event and real-time execution.
[8] Building effective agents
Anthropic · Engineering article · 2024.
[9] WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment
Hao Tang, Darren Key, and Kevin Ellis · Research paper · 2024.
[10] SOTOPIA-π: Interactive Learning of Socially Intelligent Language Agents
Ruiyi Wang et al. · Research paper · 2024.
[11] Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions
Rui Wang et al. · Research paper · 2019.
[12] Building Physical RSI Beyond the Known World
MirroS · Research position · 2026.
[13] GR00T N1.5: An Improved Open Foundation Model for Generalist Humanoid Robots
NVIDIA GEAR · Research page · 2025.
[14] Gemini Robotics
Google DeepMind · Model page · 2026.
[15] GPT-6 Astra: A new generation of intelligence
OpenAI · Product and research release · 2026.
[16] SceneSmith: AI agents create virtual playgrounds for robot training
MIT CSAIL / Toyota Research Institute · Research coverage · 2026.
[17] Agent as Policy for Robotic Manipulation
Mengzhao Jia et al. · Research paper · 2026.
[18] Claude plays robotics
Anthropic · Research article · 2026.