{"@context":"https://schema.org","@type":"Organization","@id":"https://sifted.network/organization/deepleap#onepager","name":"DeepLeap 深度跃迁 — Action Foundation Models for Robots","tagline":"Action foundation model team looking for robot OEMs, system integrators and industrial end-users to co-deploy manipulation pilots in real production lines.","description":"## What we do\n\nDeepLeap (深度跃迁) builds the action foundation model layer for embodied AI. Instead of writing a control program for every robot and every task, we train one model that learns manipulation from real-world demonstrations and generalises across tasks, scenes and robot bodies.\n\n## Technology\n\n-   **DELE-w0.5** — a world-representation route: the model plans directly in action space without generating future video frames, keeping inference fast enough for real-time control (87.5 ms median latency on an RTX 4090).\n    \n-   **Cross-task and cross-embodiment** — the same model adapts to different robot bodies, so a new scene does not require a new model.\n    \n\n- **Real-world learning loop** — UMI-based capture → data platform (curation, quality scoring, recipe, versioning) → training and evaluation → deployment feedback. Failure cases and human interventions flow back automatically as next-round training data.\n\n## Traction\n\n- Validated in **640 real-robot comparison trials** across 4 long-horizon manipulation tasks and 8 policies.\n\n- **62.5%** full-task success rate, **2.1×** the strongest baseline; **81.3%** stage-level progress.\n\n- Pilot conversations underway with energy, manufacturing and consumer-electronics groups in China.\n\n## What we are looking for\n\n- Robot OEMs and integrators who want a general manipulation model on top of their hardware.\n\n- Industrial end-users with repeatable, high-value manipulation tasks for a first pilot.\n\n- Channel and distribution partners for overseas markets.\n\n## Contact\n\n- Email: \\[juana@deepleap-x.com\\]\n\n\\- Website: \\[deepleap-x.com\\]","about":"DeepLeap builds DELE, an action foundation model that lets robots perform long-horizon manipulation tasks without task-specific programming. Founded in 2025 in Shenzhen by engineers from Huawei, UBTECH, BYD, Insta360, Li Auto and SiCarrier, the team covers model training, inference optimisation, data collection and on-site deployment. What sets us apart is the real-world learning loop: every deployment returns failure cases and human interventions as training data, so the model compounds instead of plateauing.","url":"https://sifted.network/organization/deepleap","website":"deepleap-x.com","knowsAbout":"Robot","location":"Shenzhen, China","dateModified":"2026-10-21T10:18:48.077Z","listings":{"deals":[],"opportunities":[],"events":[]},"inquiry":{"enabled":true,"endpoint":"https://sifted.network/api/organizations/deepleap/inquire","allowAnonymous":true}}