Delta Intelligence launches Δ₀ — whole-body humanoid model with latent world-action brain, 69-DoF controller, and 10,000+ h egocentric panoramic pretraining

TL;DR

Delta Intelligence (UCLA-PhD-founded, six funding rounds and ~500M yuan raised in ~six months per external coverage) publicly launched Δ₀, their native humanoid foundation model, alongside the Delta D1 head-mounted panoramic full-body data-capture device. Δ₀ pairs a latent world-action model (brain) with a learned 69-DoF whole-body controller (cerebellum), pretrains on 10,000+ hours of paired egocentric observations and whole-body motion in a shared cross-embodiment action space, and is demoed on six home tasks (record player, bed-making, toy pickup, dishwasher, step trash can, sofa sitting) with whole-body coordination and recovery from missed grasps and interference — all without teleoperation or manual resets. High-fidelity real-to-sim reconstruction is used to close the loop: sim vs real dishwasher success tracks the same trend with training-data growth (86% vs 89% at full data), and real-world RL with human-in-the-loop corrections lifts dishwasher success from 20% (4/20) to 65% (13/20).

Key claims

• Δ₀ factors into a latent world-action brain that models visual futures and actions, plus a learned 69-DoF whole-body controller that maintains coordination while accepting delta-action inputs so operators can correct motion during execution [tweet thread part 2].
• Whole-body coordinated behavior across all six demoed home tasks (record player, bed-making, toy pickup, dishwasher, step trash can, sofa sitting), with recovery demos handling missed grasps and interference without teleoperation or manual resets [tweet thread part 1].
• Pretrained on 10,000+ hours of paired egocentric observations and whole-body motion, with a shared action space across embodiments — the retargeting story is baked into pretraining rather than added downstream [tweet thread part 3].
• Delta D1, a “panoramic full-body data capture device” (external coverage: head-mounted), is released as the productionized data-collection substrate — the scale-of-embodiment-free-human-data lever for their pipeline [tweet thread part 3].
• High-fidelity real-to-sim reconstruction captures appearance, geometry, and physics closely enough that on the dishwasher task, sim and real success track the same trend as training data grows: 86% sim vs 89% real at full data [tweet thread part 4].
• Real-world RL post-training with autonomous attempts + targeted human-in-the-loop corrections lifts dishwasher success 20% (4/20) → 65% (13/20) in a mirrored kitchen [tweet thread part 5].

Method

Announcement thread only — no arXiv paper, technical report, or open weights at filing time. Architecture is described as a two-tier “brain + cerebellum”: the brain is a latent world-action model (jointly modeling visual futures and delta-action-conditioned control inputs), and the cerebellum is a learned 69-DoF whole-body controller that consumes the brain’s outputs and enforces whole-body coordination. External Chinese-language coverage (10jqka, CSDN, 36kr, BigGo) adds a “force-position hybrid control” third layer and describes a “native 3D world engine” as the substrate, with training on native humanoid-robot data captured through the Delta D1 device. Data collection is designed to work without a physical robot (via D1); training data + policy transfer across humanoid embodiments through the shared action space. Sim-to-real is closed by physically-accurate reconstruction (appearance + geometry + physics) rather than pixel-space rendering only. Real-world RL uses autonomous rollouts + human-in-the-loop corrections as the improvement signal.

Results

• Six home tasks demonstrated with whole-body coordination: record player, bed-making, toy pickup, dishwasher, step trash can, sofa sitting [tweet thread part 1].
• Dishwasher, sim vs real scaling parity: 86% (sim) vs 89% (real) at full training data — sim success tracks real success across scale [tweet thread part 4].
• Dishwasher, real-world RL post-training: 20% → 65% success rate (4/20 → 13/20) after RL with targeted human-in-the-loop guidance [tweet thread part 5].
• No head-to-head with π0.5 / GR00T / Isaac 0.5 / other filed humanoid VLAs disclosed; no per-task success rates for the other five home tasks; no backbone size, action-head architecture, or training compute reported.

Why it’s interesting

Δ₀ adds a whole-body humanoid + real-to-sim scaling parity + real-world-RL closed loop cell to the recipe-lever board on VLA Models that no filed system currently occupies at this scale — the closest neighbors are Gemini Robotics 2 brings whole body intelligence to robots Gemini Robotics 2 (whole-body humanoid VLA, but closed and no scaling numbers), HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own HomeBody (whole-body Unitree G1 via LLM-as-controller with five typed skills, no learned 69-DoF controller), and Isaac 0.5: An Open-Weight Embodied Foundation Model Isaac 0.5 (Perceptron’s 210× teleop-reduction scaling with general video, but bimanual not whole-body). The 86% sim vs 89% real parity at matched data scale is the sharpest published sim-to-real fidelity number on this concept page and directly answers the “does R2S2R actually work for policy scaling” question Building Worlds That Train Robots — Real-to-Sim-to-Real (R2S2R) as a Scalable Engine for Training and Evaluating Robot Policies raised. The 20% → 65% RL lift with human-in-the-loop corrections slots directly into the intervention-scaling thesis of MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory / Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact / Robo Robotics launch — Robo-T bimanual humanoid, sub-$10/hour with Roboport teleoperation platform (Kyle Noble) Robo Robotics, but scaled to a whole-body humanoid rather than arm-only cells. The D1 head-mounted panoramic capture device sits next to Ego-OSCAR: Egocentric Open source Stereo CAptuRe System Ego-OSCAR and Lightwheel AI open-sources EgoSuite-Open100K — 100K-hour egocentric human dataset with hand and body pose EgoSuite as first-party data-collection hardware, and the “shared action space across embodiments” claim generalizes the Human-to-Robot Retargeting problem into a pretraining-time bake-in rather than a downstream retargeting step.

See also

• VLA Models — Δ₀ stakes out the whole-body-humanoid + latent-world-action-brain + 69-DoF-controller cell of the recipe-lever board
• World Foundation Models — the “latent world-action model” brain is the WFM half of the Δ₀ stack
• Human-to-Robot Retargeting — 10,000+ h with a shared cross-embodiment action space bakes retargeting into pretraining
• Synthetic Training Data — sim vs real dishwasher parity (86% vs 89%) is a scaling datapoint for R2S2R pipelines
• Reasoning RL — 20% → 65% real-world RL with human-in-the-loop corrections
• Pose Estimation and Motion Capture — Delta D1 head-mounted panoramic capture is the whole-body-motion labeling substrate
• Gemini Robotics 2 brings whole body intelligence to robots — closest closed-flagship neighbor (whole-body humanoid VLA)
• HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own — whole-body G1 via LLM-as-controller alternative to Δ₀’s learned 69-DoF controller
• Isaac 0.5: An Open-Weight Embodied Foundation Model — clean scaling numbers on video pretraining, bimanual rather than whole-body
• Building Worlds That Train Robots — Real-to-Sim-to-Real (R2S2R) as a Scalable Engine for Training and Evaluating Robot Policies — R2S2R framing that Δ₀’s sim-vs-real parity directly bears on
• MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) — human-in-the-loop DAgger intervention-scaling thesis in an industrial context