Delta-0 (Δ₀): A New Chapter in Humanoid Intelligence

Introduction

The humanoid form offers the possibility of working wherever people do. Humanoid intelligence is what turns that possibility into useful and reliable action.

Introducing Δ₀, Delta Intelligence’s humanoid foundation model (HFM) for whole-body loco-manipulation. Δ₀ learns from human motion and robot interaction to coordinate perception, locomotion, and manipulation through a learned whole-body controller.

Our homes, workplaces, tools, and appliances are organized around human reach, movement, and dexterity. A humanoid body is suited to these environments and offers a connection to human experience: how people move and handle objects across tasks and settings. Robot interaction grounds that experience in the robot’s own capabilities.

Whole-body loco-manipulation couples movement with object interaction while maintaining balance and managing contact. In long-horizon tasks, every action changes the conditions for what comes next. Completing the task means preserving balance, maintaining useful grasps, and arriving in positions that allow the robot to continue.

The household demonstration brings this pursuit into focus: making a bed, picking up objects from the ground, opening a dishwasher, and operating a step trash can each require coordinated movement and manipulation, connected across a sequence of everyday work.

Loco-manipulation: the most challenging unsolved problem

At Delta Intelligence, we view general, reliable whole-body loco-manipulation as the core problem of humanoid intelligence—and the most challenging unsolved problem on the path to useful general-purpose robots. It brings locomotion, whole-body control, and manipulation together: moving through an environment, maintaining balance under contact, and acting on objects. Each capability continuously constrains the others.

The robot must control a moving body while perceiving and manipulating a changing world. A humanoid has a floating base: its position and orientation change as it moves, and its balance depends on environmental contacts. Stepping and posture adjustments affect both hand position and, with body-mounted cameras, visual perspective. Dexterous manipulation makes this coupling especially demanding. Small body movements can disturb hand–object alignment, so precise interaction requires coordinated control of the hands, torso, and legs.

Long-horizon tasks compound these difficulties. Carrying a cup to a dishwasher requires a grasp that survives transport, walking that keeps the cup stable, and an arrival stance that supports placement. Each action determines the starting conditions for the next. Small errors can accumulate, and a successful intermediate action may leave the robot poorly positioned to continue. Useful execution must preserve balance and control across these transitions while maintaining a practical pace.

The same task must succeed across changing viewpoints, body configurations, support contacts, and object poses. Learning requires experience that covers these variations and a controller capable of executing the resulting actions. Human activity provides diverse coordination patterns; robot demonstrations ground them in the robot’s capabilities. Corrections and physical interaction help address deviations, while repeated sequence evaluation—through simulation and hardware—tests whether improvements hold across the complete task.

Δ₀: six steps toward humanoid intelligence

Long-horizon loco-manipulation demands coordination, practical speed, transferable experience, executable control, corrective learning, and scalable evaluation together. Δ₀ brings these requirements into one learning system.

Humanoid intelligence in everyday tasks

Everyday life involves a wide range of physical tasks: putting on a record, making the bed, picking up a toy, opening the dishwasher, using a step trash can, and sitting down on a sofa. These familiar activities require careful handling, coordinated movement, and balance as the body interacts with objects and furniture. Across the six tasks, Δ₀ adjusts its reach, posture, and support to meet these different demands, bringing whole-body coordination into practical household activity.

Two halves of one mind: co-designing the brain and controller

Δ₀ pairs a brain—a latent world-action model—with a learned whole-body controller. The brain predicts future latent world states and produces motion commands; the controller turns those motion commands into coordinated physical action.

We co-design the brain and controller so that the brain can express the motions a task requires and the controller can reliably execute them.

The learned whole-body controller: a foundation for humanoid intelligence

The human cerebellum is estimated to contain about 80% of the brain’s neurons (Azevedo et al. (2009)). For humanoid robots, the analogy motivates a design principle: The whole-body controller should be learned, continually improved, and scaled alongside the brain.

Δ₀’s whole-body controller learns from diverse motion data to coordinate 69 degrees of freedom. It turns the brain’s motion commands into joint-level movement while maintaining balance and responding to contact.

The brain: a latent world-action model

Jointly learning future visual states and motion commands connects the robot’s actions with how the scene is expected to change. As shown in the architecture figure above, the mixture-of-transformers (MoT) architecture gives visual dynamics and whole-body motion dedicated modeling capacity while sharing information from visual observations, language, and the robot’s own state.

The architecture comprises three specialized branches. The vision-language branch processes current multi-view observations and the task instruction. The vision branch predicts future visual states in a semantic feature space provided by DINO, while the action branch models whole-body motion conditioned on proprioception. Each branch retains its own attention projections and feed-forward layers, while shared multimodal self-attention allows information to flow across the three streams.

We train the model with four complementary modes by varying the conditioning inputs and denoising targets. Forward dynamics predicts future visual features given actions, while inverse dynamics predicts actions given future visual features. Visual planning predicts future visual features without action conditioning, while policy-only training predicts actions without conditioning on future visual targets. Current observations and task context provide the common grounding across these modes. Together, these objectives train a single model to capture the relationships between observations, visual futures, and actions.

Long-horizon execution through stage-by-stage instructions. Δ₀ carries out extended tasks through stage-level instructions supplied by a higher-level model, agent, or a human. Connecting tasks is itself a whole-body challenge: completing one task does not necessarily leave the robot ready for the next. The instructions specify the next objective; Δ₀ must learn how to reposition its body, adjust its stance, and preserve or release a grasp so it can reach the next object and interact with it while staying balanced. Our focus is on these physical transitions, rather than a dedicated high-level task planner.

Scaling humanoid intelligence requires expanding what robots can learn, how thoroughly their capabilities can be evaluated, and how they improve through interaction. Human activity provides a rich source of coordinated behavior. Reconstructed environments make repeated evaluation across tasks and conditions more practical. Real-world experience provides feedback for refining skills and learning from failures. These three pathways—data, evaluation, and experience—connect broader learning with increasingly reliable physical execution. Δ₀ brings them together through human-data pretraining, real-to-sim-to-real evaluation, and real-world reinforcement learning.

Scaling data: learning from human data

Human activity offers a rich source of whole-body skills: people coordinate their hands, posture, and movement to interact with the world. The similarity between human and humanoid body structure makes this experience a useful starting point for learning both how to move and how to act in a scene.

We train the controller on large-scale high-quality human motion data, expanding its repertoire of coordinated, executable movements. We measure this foundation through tracking accuracy as the motion dataset grows: how faithfully the controller executes a requested movement. This measures execution quality, while the brain’s success is measured by whether its actions accomplish a task.

The brain pretraining utilizes over 10,000 hours of paired human data. Paired egocentric observations and whole-body motion connect what a person sees with the actions they take. These examples capture how people establish support and reach—stepping closer, lowering the torso, or changing posture—as they handle objects. Pretraining on this experience exposes the policy to varied scenes, intermediate postures, and object configurations.

We propose a shared action representation for Δ₀ that enables the brain to learn from different data sources and embodiments. Its 154 dimensions cover arm and hand actions, root commands, and motion commands, padded to a common 180-dimensional layout. Each data type contributes the channels it provides.

The examples below showcase our own collected egocentric data, pairing video with continuous whole-body motion and detailed hand reconstruction. The synchronized views connect the person’s perspective with body movement and finger motion during everyday activities.

We evaluate controller scaling by measuring high-precision tracking coverage on everyday-task motions at increasing motion-data budgets. As motion data scales, coverage rises steadily, reflecting clear gains in controller capability.

We evaluate policy scaling by measuring task success at increasing pretraining data budgets. In our internal evaluations, task success improves with more pretraining data across three task categories: basic loco-manipulation that a mobile manipulator could also perform; bimanual tasks requiring fine manipulation; and whole-body tasks involving contact beyond the hands, such as closing a cabinet with a leg.

We also examine how the policy responds to unfamiliar conditions and disruptions through the following demonstrations. While loading a cup into the dishwasher, Δ₀ operates under shifting colored light it never saw in training. The paired views show the scene and the robot’s onboard observations.

The recovery clips show a missed grasp, a new cup placed on the counter after the first is loaded, and a person pushing the trash can lid shut. In these demonstrations, Δ₀ responds autonomously and completes the task without teleoperation or a manual reset.

These demonstrations show adaptation within individual tasks. Establishing how consistently a policy handles such variations requires systematic evaluation across repeated trials and conditions.

Scaling evaluation: closing the real-to-sim-to-real loop

Scaling evaluation is a bottleneck to accelerating humanoid model development. As training produces more candidate models, progress depends on quickly determining which changes improve performance and where failures remain. Testing every checkpoint on a physical humanoid is slow, costly, and limited in coverage, delaying the feedback needed to guide the next training cycle. Real-to-sim-to-real evaluation addresses this bottleneck by reconstructing real tasks in simulation, comparing policies through repeatable trials, and returning selected candidates to hardware for validation. This shortens the feedback cycle between training, evaluation, and model refinement.

We reconstruct deployment environments and task-relevant objects as calibrated simulation assets. Our pipeline combines 3D reconstruction methods with agentic frameworks built on cutting-edge large language models (LLMs) such as GPT-6 Astra, which automatically turn video captures of a scene into simulation-ready assets with high visual, geometric, and physical fidelity, covering object geometry, materials, articulated components and hinges, contact surfaces, friction, the robot and its sensors, and the support beneath it. Editable properties allow controlled changes to lighting, appearance, friction, and object configuration while retaining the structure of the real task.

Within these environments, Δ₀ runs autonomous rollouts that connect navigation, manipulation, and retries across task stages. Repeated evaluation on the same task suite and controlled variations supports checkpoint comparison and reveals failures between stages that isolated trials may miss.

On the dishwasher task, simulated and real-world success rates follow the same overall improvement trend as training data increases. This provides an early signal for comparing training changes before hardware validation.

Selected policies return to the real robot to test performance under physical contact, sensing, actuation, and balance demands. Hardware trials also check whether the relative ordering of candidates in simulation holds in reality. Discrepancies inform the next iteration of reconstruction and evaluation, closing the loop between repeatable simulated tests and physical execution.

Scaling experience: humanoid real-world RL post-training

Real-world reinforcement learning extends policy training beyond demonstrations by learning from autonomous attempts and targeted human guidance. Δ₀’s delta-action controller gives policy outputs and human-in-the-loop (HIL) corrections a shared execution interface.

A value model estimates task progress from visual observations, recent history, and robot state. During RL post-training, action chunks with larger gains in predicted progress receive a positive conditioning signal; those with smaller gains or setbacks receive a negative signal.

Human corrections supply recovery examples where the autonomous policy struggles. HIL correction chunks receive a positive conditioning signal directly, allowing targeted guidance to contribute alongside the model’s assessment of autonomous attempts.

What comes in the future

Scaling humanoid intelligence. A central question is how larger models and more varied training data translate into broader humanoid capabilities. Human activity, robot demonstrations, and interaction with the world offer complementary sources of experience, with the potential to support richer coordination and more adaptable behavior.

Toward reliable deployment. These capabilities become useful when robots can perform them dependably in everyday settings. Long-horizon tasks and changing environments bring questions of consistency, recovery, and practical execution speed into focus. Closing the distance between demonstrated skill and dependable everyday behavior remains central to the broader pursuit of humanoid intelligence.