Symbiosis Robotics released a new demo showing a humanoid robot driving a go-kart, and introduced Direct Perception Control (DPC), which eliminates intermediate motion representations and directly generates joint actions from visual and task information, aiming to bridge the gap between perception and whole-body control in hierarchical systems. The team compiled 15,010 hours of embodied data for training.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Symbiosis Robotics unveils DPC technology, which eliminates intermediate motion representations, allowing robots to generate joint actions directly from vision and body state, as demonstrated in a go-kart driving demo.
Symbiosis Robotics has released a new demo where a robot drives a go-kart and performs drifts, but the focus is on the underlying Direct Perception Control (DPC) technology. DPC eliminates the intermediate motion representation that requires a standalone Whole-Body Tracker, allowing the model to combine visual input, task, and body feedback to directly compute joint and hand targets.
Traditional hierarchical architectures like Figure Helix and Gemini Robotics 2 handle semantic reasoning and whole-body control separately, whereas DPC explores joint optimization of the cognitive model and motion expert in unified training, enabling task understanding and whole-body control to adapt to each other.
Internally, DPC retains a Perception Expert and a Motion Expert, which exchange information via Symbiotic Attention, and are trained under the same action objective, asynchronously updating visual results while continuously reading proprioceptive feedback.
Directly outputting Joint Targets requires converting data from different sources into the same 'joint language.' DPC converts human videos and various robot data into G1-executable, time-aligned joint actions, which are then used uniformly for training.
The updated dataset totals 15,010 hours: 6,781 hours of Human Ego data, 4,024 hours of Armed Robot data, 3,660 hours of Wheeled Humanoid data, and 545 hours of Bipedal Humanoid data.
After conversion, visual input, task language, body state, action history, and subsequent joint actions are placed on the same timeline, providing DPC with unified Joint Target supervision.
After removing the intermediate interface, the DPC policy must handle drift during real-world execution. It first initializes via Offline BC, then enters the second stage of DriftDistill training.
During online rollouts, the Student encounters states that deviate from training; a Frozen Teacher provides recovery actions for these states. During training, online rollouts and offline demonstrations are used together to update the Student.
Through repeated rollouts, the Student gradually learns to handle deployment states and recovery paths, allowing it to continue tasks after deviation.
The official page also showcases mobile pick-and-place and whole-body manipulation under constrained poses. The robot walks to a post, bends down to pick up a target with one hand, then turns and moves to a designated location to place it.
Sitting in the go-kart pushes the capability into a more constrained space; once on the track, the changing visual input enters the control loop. These demos illustrate the same point: the robot directly computes the next joint action after perceiving environmental changes.
But a drift video does not mean the robot has learned general driving. DPC's demonstration focuses on whether task understanding and whole-body control can truly be connected under a unified training objective.
This article is based on a report by Machine Intelligence (机器之心) on Symbiosis Robotics' official technical page; all technical details and figures come from that source. Some comparative information (e.g., Figure Helix, Gemini Robotics 2) is background reference and has not been independently verified.
DPC eliminates intermediate motion representations to let perception directly drive joint actions, trained on 15,010 hours of multi-source data, showcasing a new direction for humanoid robot control, but its generality still requires further validation.
Primary report
Primary source