Key points
- Delta Intelligence releases Delta-0, a humanoid foundation model controlling 69 degrees of freedom on a Unitree G1.
- Delta-0 trained on 10,000 hours of data toward a 100,000-hour goal, and its dishwasher task scored 89 in simulation versus 86 in reality.
- Siyuan says real-world RL is needed to move humanoid success from 60-70% to 99% for deployment.
Delta-0 runs a humanoid's whole body from one model
Delta Intelligence is betting that humanoid intelligence comes from co-designing the brain and the whole-body controller, not from bolting a large model onto an existing robot. Its first model, Delta-0, drives 69 degrees of freedom on a Unitree G1 and runs locomotion and manipulation from one policy, trained on full-body human motion capture because nearly all egocentric data stops at the waist. The demo is three minutes of uncut housework at near-human speed. The real claim is about the last mile, where real-world RL and human intervention push 60 percent success toward the 99 percent deployment needs.
Takeaways
- One model drives a whole humanoid body. Delta-0 controls 69 degrees of freedom on a Unitree G1 with two WUJI hands, against 29 on the base robot. A single model handles locomotion plus upper and lower body interaction, with no SLAM navigation.
- Human data is the richest training source. Roughly 90 to 99 percent of egocentric recordings cover only the upper body, so Delta Intelligence added a fisheye camera aimed at the body and plans foot trackers. Training used about 10,000 hours, with 100,000 as the goal.
- Real-to-sim-to-real pipeline anchors the work. The team rebuilt the demo room in 3D and used Astra for automatic physical system identification, with an agent handling about 90 percent of simulator parameter tuning. The dishwasher task scored 89 in simulation and 86 in reality.
- Real-world RL closes the last gap. Siyuan says the field lacks recipes for turning 60 or 70 percent success into 99 percent. Human-in-the-loop learning lifted an out-of-distribution task from about 20 to 65 percent, and real-world RL is necessary where no 3D replica exists.
- Humans can intervene during any rollout. The delta-action controller lets an operator take over mid-inference through a PICO teleoperation rig when the robot errs. Siyuan says Delta Intelligence is the first to build a delta-action control system for humanoids.
- Active vision was dropped from the design. Active vision gave no significant improvement, so the robot carries an Insta360 X5 for a panoramic view while body movement compensates for camera motion. Neck and body motion are entangled in egocentric data, which drove the full-body capture system.
What comes next
- Siyuan predicts a turning point within a year between what humanoids can do and where they can be deployed, with real-world RL as the bridge.
- Data scaling runs from about 10,000 training hours toward 100,000.
- Foot trackers will fuse hardware tracking with model estimation to handle occlusion.
- Contact and tactile sensors come after the team explores the limits of visual input, since tactile sensing proved unstable and marginal where vision is plentiful.