Latest / Embodied AI 101 / Delta-0: Why Humanoid Foundation Models Need Correctable Control
Transcript
- Delta-0: Why Humanoid Foundation Models Need Correctable Control. Delta-0 (Δ₀): A New Chapter in Humanoid Intelligence • Delta Intelligence Team • Delta Intelligence • Technical blog • 2026. Cold Open. A humanoid opens a dishwasher by pulling with one hand and bracing with the other. Delta Intelligence makes this coordinated contact a centerpiece of its Delta-0 household demonstration. Now imagine the reaching hand approaches the handle at the wrong angle. An operator sees exactly what needs changing: move the hand sideways, then continue pulling. How should that person intervene while the robot is bent over, touching furniture, and managing its balance?. You want to correct the reach. You do not want to become responsible for every joint in the robot. That is the interesting question underneath this release. A humanoid learning system needs a connection between autonomous attempts and useful human corrections. What should that connection look like? Which decisions should the person supply, and which should remain with the controller?. Delta-0 puts that boundary at the center of the design. Today on Embodied AI 101, we are looking at what happens between a policy making a mistake and that mistake becoming useful training experience. The Idea. On hardware, Delta-0 controls a Unitree G1 with two WUJI hands, totaling 69 degrees of freedom. Its learned whole-body controller is trained through reinforcement learning in simulation. Above that controller sits a latent world-action model. It combines camera observations, language, and robot state through specialized vision-language, future-vision, and action branches. Those branches exchange information through multimodal attention. Future observations are represented as DINO features, while the action branch produces motion commands for the controller. Training alternates among four complementary questions: Which action explains a future observation? What future follows an action? What future should be anticipated without a supplied action? And what action should be taken directly? These correspond to inverse dynamics, forward dynamics, visual planning, and policy learning. Think about the advantage of that arrangement. An observation, an action, and its consequence can provide several kinds of supervision. The learning problem becomes richer than matching a camera frame to a demonstrated command. The model is also asked to connect commands with anticipated changes in the scene. Delta reports pretraining on more than 10,000 hours of paired human observations and motion, using a shared action representation across embodiments. The intended connection is between what a person sees and how they coordinate their body to act. The company’s public OmniEgo data provides a concrete view of its capture approach: synchronized egocentric cameras, headset motion, and whole-body skeleton information. The learning hypothesis is straightforward. A useful demonstration should expose the preparation for manipulation as well as the final hand movement. Stepping closer, lowering the torso, and establishing support are part of the behavior you want the robot to acquire. Then comes the intervention interface. The authors describe pausing the policy to supply relative end-effector corrections while retaining the whole-body controller. An operator can redirect a hand without first matching the robot’s complete posture. That changes the meaning of taking over. Imagine correcting a reach while sitting at a desk, even though the robot is crouching. The person communicates the desired adjustment; the controller remains responsible for realizing it through the body. For learning, the attractive consequence is more specific supervision. Instead of requiring another pristine demonstration from the beginning, the interface can support a correction at the point where the autonomous behavior departed from what was useful. The training example starts from the robot’s actual situation. Long-horizon execution uses stage-level instructions supplied by a higher-level model, agent, or person. Delta-0 handles the physical transitions between those objectives. That division makes the co-design argument concrete. Choosing the next objective, producing an executable motion, and correcting that motion are connected decisions. The action interface determines how information passes between them. Designing it alongside the controller gives both autonomous behavior and human guidance a common route to the motors. The Result. The headline result is dishwasher success increasing from 20% to 65% after real-world post-training with targeted human corrections. Each reported evaluation used twenty trials. The experiment took place in a mirrored kitchen layout, and the improvement followed collection of corrective experience in that layout. The post-training recipe uses a value model to estimate task progress. Autonomous action chunks receive conditioning signals based on estimated progress, while human correction chunks receive positive conditioning directly. The important connection is between the control interface and the learning signal. The operator is helping at states encountered during the robot’s own execution. That is exactly where a recovery example can be valuable: after a reach has drifted, after a grasp has gone wrong, or after the robot has arrived somewhere awkward. For an experimentalist, this suggests a productive unit of progress: an autonomous attempt that can be corrected and reused. The interface matters because it determines whether that experience is straightforward to collect. The success-rate change gives the overall adaptation loop a measurable outcome. There is also supporting evidence at the controller level. Delta reports a winning tracking comparison against HEFT, MimicLite, SONIC, and ScaleBFM on its everyday-task motions. It uses the public Motion Tracking Leaderboard’s evaluation workflow. That benchmark ecosystem distinguishes measures such as body-position accuracy, global root tracking, wrist tracking, and motion progress. These measure different things. For manipulation, reproducing a plausible posture and arriving at the intended interaction location are both useful properties. A controller evaluation should make that distinction visible. The scaling experiments report improving tracking coverage as controller motion data grows, alongside improving task success as policy pretraining data increases. The architectural story has corresponding measurements: execution capability on one side, task accomplishment on the other. Delta also reports that simulated and real dishwasher performance follow the same improvement trend as training data increases. Its reconstructed environment is intended to support repeated policy evaluation before hardware validation. The engineering opportunity is faster experimental selection. A useful simulator can help decide which candidate deserves the next physical run, which transition deserves closer inspection, and which training change is worth pursuing. Finally, the authors present extended household execution at native playback speed, describing its pace as near-human. The quantitative adaptation experiment supplies the clearest evidence for the episode’s central claim: corrective interaction can feed back into improved humanoid task performance. The Biggest Weakness. The evaluation does not establish broad transfer—the central promise attached to calling this a foundation model. The mirrored-kitchen result comes after collecting new experience in the mirrored kitchen. It demonstrates adaptation to that setting. The report does not present a held-out task-and-environment benchmark that measures how broadly the resulting capabilities transfer. This is a distinction between experimental questions. One question is whether a capable humanoid system can improve on a particular task through additional interaction and human assistance. Another is whether its pretrained knowledge makes unfamiliar tasks and environments substantially easier to learn. A foundation model does not have to solve everything zero-shot. Faster adaptation can be compelling evidence of a reusable foundation. The key is measuring that benefit across genuinely held-out situations under a controlled adaptation budget. The decisive experiment would reserve unfamiliar rooms, appliances, and task combinations before development. Then compare adaptation curves while holding the controller, interaction budget, and available human assistance constant. Measure autonomous completion of the entire requested sequence, together with the operator effort required to reach that performance. The question would become concrete: does this pretrained system bring reusable coordination knowledge, or does each new setting demand a similar amount of task-specific teaching?. That matters for practitioners deciding where to invest. A broadly reusable model changes the economics of onboarding new work. A system that learns each installation separately creates a different operational commitment. Delta-0’s existing results support a serious systems-learning story. The foundation-model thesis needs an additional result: a measured advantage that survives changes in the task and environment, with the cost of adaptation made explicit. That is the strongest unresolved question raised by this release. Verdict. VLA teams moving onto bipedal platforms should pay attention. So should teleoperation engineers and researchers building real-world reinforcement-learning loops. The most reusable idea is to treat the action interface as part of the learning method. It determines what the policy can request, what a person can correct, and what experience the system can collect. Those are modeling decisions as much as controller decisions. Two related research threads make the contribution easier to place. NVIDIA’s SONIC scales human motion tracking and connects teleoperation and VLA-driven behavior through a shared control interface. Its demonstrations include coordinated whole-body loco-manipulation. Berkeley’s HIL-SERL places human interventions inside a real-robot reinforcement-learning process, combining demonstrations, autonomous experience, and corrective guidance. Read Delta-0 at that intersection: a learned whole-body execution layer designed to support correction-driven improvement. The systems contribution is the deliberate connection between those agendas. A useful companion is Zero-HLM, an October 2026 paper that builds reconstructed environments and simulation-generated demonstrations for humanoid loco-manipulation. It trains policies that execute through SONIC on a real G1. That work offers another route for producing experience: construct and validate whole-body interactions in simulation, then transfer the learned behavior to hardware. For system builders, the attractive research direction is combining these sources of experience deliberately. Human motion can supply coordination examples. Simulation can expand interaction coverage. Corrective real-world experience can teach recovery from the situations the deployed policy actually encounters. There are concrete places to start studying that design space. SONIC and HIL-SERL publish implementation resources, while Delta’s OmniEgo release exposes synchronized visual and body-motion data. My takeaway is simple: for humanoid learning, a correctable controller is part of the model. Build the route from autonomous action to useful correction with the same care you give the route from observations to actions. Outro. Return to the dishwasher handle. The operator knows what should change. The robot needs to preserve everything that should stay coordinated. A good intervention interface makes that division of responsibility clear. That is a powerful design question to carry into your own system: what is the smallest useful correction a person can communicate, and how much physical responsibility can the robot retain while accepting it?. Watch how these interfaces evolve from rescuing individual attempts to teaching new skills. The next useful training example may begin halfway through an imperfect rollout, with a small correction that the whole body knows how to execute.