Source Future Unveils First Whole-Body Intelligence Robot Demonstrations

The demo reported by Machine Heart begins with a simple scenario: a doll is placed too high, the robot cannot reach it, so it brings over a step stool, stands on it, and takes the doll down. It looks like just "stepping up a bit," but moving the stool, stepping onto a height, and reaching for the object—each step changes the body's support conditions.

The report uses a step at the kitchen entrance as an entry point: to a human it is one step, but to a robot it is a round of support transition—one foot lifts first, the other braces the body, and after landing the center of gravity is shifted over. With the two feet not at the same height, the body cannot continue copying its posture from flat ground.

YuanCe's self-developed S0 whole-body control system coordinates foot trajectories, torso posture and center-of-gravity shifts. According to the report, stepping into the kitchen and bending down to pick up objects on the step both test body control when support conditions change. Reaching the doll after climbing is a further benefit; what must first be solved is that with the ground beneath changed, the body can still act stably.

It should be noted that all of the above are demo descriptions of a vendor-release nature; the source gives no quantitative performance metrics, nor does it state whether the demos were completed continuously and autonomously, or whether they included human intervention or editing.

The demo does not stop at climbing to reach objects. A water bottle lies fallen by the robot's feet; the robot bends its knees, leans down to grab it, and rises with the object. The report emphasizes that when body height changes, the target in hand must not lose accuracy along with it.

In another demo, a bicycle lies fallen by the roadside; the robot bends down to grab the frame and then lifts it up. As the frame leaves the ground, both the load in hand and the body's support are changing. Carrying a box is the same: from hugging the box to placing it on a shelf, the weight transfers from both hands to the shelf, and the arms, waist and legs must coordinate to complete this "handover."

The camera cuts back to the kitchen, where the target shrinks to a single button on a rice cooker. This time no large movement is needed, but if the body sways, the fingertip may miss the contact position. The report's summary is: the motions differ in size, yet the requirements are linked—the body must have sufficient range of motion, and key operating positions must not lose accuracy as the body moves.

These tasks are all artificial or controlled scenarios; the source does not disclose the model sizes, training data volumes or compute investment of S0 and S1.

After the items on the dining table are cleared, G1 does not stop, but picks up a rag and continues wiping the tabletop. The report points out that in pick-and-place the hand must pick up and release, while in wiping the tool must move pressed against the tabletop—both look like "clearing the table," but to the robot they are two different modes of contact.

The dishwasher scenario is denser: the door rotates around a hinge, the dish rack travels along a slide rail, and the dishes must be aligned into position. According to the report, T2 cannot handle the whole process from start to finish with the same pulling motion.

Behind this is a division of labor: S1 organizes task actions based on on-site state, while S0 handles whole-body coordination and stable execution. The previous step changes the object's position and also changes where the next step begins. The report's observation is that finishing a single grasp and finishing a whole task are separated by a series of transitions.

This division-of-labor description comes from the vendor's release; the source provides no verifiable data such as task success rates or completion times.

In front of the refrigerator, G1 and T2 take part in the same organizing task. Together with the dining table and dishwasher demos on the two embodiments, the report says this technical route has already entered real operation on different bodies.

The report also points out the difficulty: body proportions, joint configurations and reach ranges all differ. The meaning of cross-embodiment lies in translating task requirements into actions each body can execute, rather than merely giving the robot a different shell.

The source provides no specific task success rates or comparison results for the cross-embodiment validation, so "has already entered real operation on different bodies" should be understood as a demonstration-level showcase, not a quantifiable proof of generalization capability.

The report uses a familiar action to illustrate the orientation of data collection: when people take something from a low cabinet, they often first shift half a step, then lower their head, bend down and reach out. Watching only the hand's trajectory easily misses why that half step matters.

EgoBody records exactly this kind of whole-body experience: the first-person view captures "what was seen," while synchronized body motion captures "how to stand, how to reach, how to coordinate." YuanCe designs tasks internally, distributes them to collection sites in different cities across the country, and then routes them back for quality inspection, bringing different scenarios, objects and skills into one data pipeline.

On the training path, S1 first learns priors from human data, then adapts to specific bodies and tasks through real-machine post-training. The report states clearly that human experience is the starting point; applying it to a robot still requires practice.

Practice is not only successful demonstrations. When execution deviates, humanoid DAgger introduces human corrections, bringing "how to adjust after going off track" into subsequent training. S0 runs through both collection and deployment: at one end it lets human demonstrations be executed on the robot, and at the other it lets the actions learned by the model truly land.

YuanCe Future was co-founded by Li Hongyang, Li Tianyu and Chen Li, focusing on a general-purpose whole-body humanoid foundation model. In July, YuanCe proposed the Whole-Body Intelligence route in a technical blog; now the footage already shows complex terrain control, autonomous tasks and cross-embodiment validation.

The report mentions that the founding team's technical accumulation spans end-to-end intelligence, robot learning, world models and large-scale mass-production autonomous driving, having led projects such as the end-to-end autonomous driving UniAD, the HKU-AgiBot ultra-large-scale embodied manipulation AgiBot World, the ByteDance version of UMI upper-body manipulation Egonomy, and Huawei's mass-production autonomous driving ADS 4.0 World Engine, among which UniAD won the CVPR 2023 Best Paper Award and AgiBot World received an IROS 2025 Best Paper Finalist Award.

The report's statement of the long-term goal is: to let experience accumulate into reusable capabilities, so that every change of job does not require teaching from scratch. The value of Whole-Body Intelligence ultimately lies not in how "human-like" the robot moves, but in how much work that originally had to be done by human hands can gradually be handed to machines to "execute."

What needs to be distinguished is: the company's founding time, founding team and paper awards are verifiable background; whether the "Whole-Body Intelligence" route can deliver on reusable capabilities currently has only demonstration-level evidence, and the source does not state whether these real-machine results have entered commercial deployment or customer scenarios.