How Six‑Dimensional Force Data Powers China’s First Full‑Perception VTLA Model

Recent breakthroughs such as Generalist AI’s GEN‑1 model, which raised robot success rates to 99% and tripled folding speed, highlight a shift in scaling laws: large‑language models rely on abundant internet data, while embodied intelligence now depends on massive real‑world interaction data.

Unlike easily replicated visual, image or video data, embodied AI requires physical interaction logs—each grasp, move, contact or failure is unique and non‑reusable. Consequently, data acquisition has become the primary bottleneck limiting progress.

Kepler Robotics announced China’s first native full‑perception force‑tactile data collection system. The platform standardizes hardware, data structures and native model integration, allowing cross‑task, cross‑scenario and even cross‑robot data reuse, turning data from a consumable into a continuously accumulating production asset.

The system drives a paradigm shift from vision‑centric imitation learning to a force‑centric, fully‑perceptive interaction model, where robots learn not only how to move but also when and how strongly to make contact.

Six‑dimensional force data comprises three linear forces (F_x, F_y, F_z) and three torques (M_x, M_y, M_z), describing push/pull and twist/torque respectively. This tactile feedback lets robots adjust actions in real time, replacing rigid pre‑programmed trajectories with adaptive, contact‑aware control.

Kepler’s dual‑path approach separates precision and scale. The “precision” path uses a bidirectional tele‑operation loop with force‑feedback exoskeletons, tactile gloves and high‑resolution fingertip sensors, achieving 99% data fidelity and millisecond‑level latency with noise below 1%.

The “scale” path adopts a UMI‑style human demonstration pipeline: operators wear tactile gloves that capture visual, joint, pressure and muscle signals, which are then mapped to robot kinematics, dramatically increasing data volume while keeping costs low.

Both paths address practical challenges: multi‑robot target mapping and strategy distillation enable data reuse across dozens of robot bodies; multi‑view camera fusion fills visual blind spots caused by occlusion or reflective surfaces.

On the modeling side, Kepler introduced a VTLA (Vision‑Touch‑Language‑Action) large model that processes multimodal inputs—RGB‑D streams, language commands, joint states and six‑dimensional force data—in a unified encoder, outputting end‑to‑end control commands.

Training leverages existing vision‑language models, augmented with simulated, real‑world and human‑video data. Evaluation now measures not only task completion but also contact stability, force accuracy and generalization to novel objects, reflecting the added tactile dimension.

In a real automotive factory line, the VTLA model performed 1,000 consecutive high‑precision assembly operations with a 99.4% success rate, a 19.4% improvement over a vision‑only baseline, and required no human intervention, dramatically reducing rework and labor costs.

Overall, Kepler’s system resolves long‑standing industry issues—misalignment between human‑collected and robot‑used data, hardware heterogeneity, and the trade‑off between data quality and quantity—by delivering a sustainable, high‑fidelity, scalable data foundation for embodied AI.