Visics, a 36Kr-featured embodied-AI startup, released Visics, a general embodied model that uses "Object Trajectory" as a unified intermediate representation. The core idea: every embodied task — from robot manipulation to autonomous driving — can be reduced to "predict the trajectory of an object over time," and a unified trajectory tokenizer can serve as a common "Token" for embodied LLMs.
The technical details: Visics's trajectory tokenizer takes raw sensor input (camera, LiDAR, IMU) and produces a "trajectory sequence" — a sequence of 6D pose vectors (position + orientation) for each object in the scene, sampled at 30 Hz. The trajectory is then fed into a Transformer-based LLM, which predicts the next-step trajectory of each object, conditioned on the task specification.
The unification value: traditional embodied models are task-specific — a manipulation model, a navigation model, a grasping model. Visics's trajectory-based approach is task-agnostic — the same model can be trained on manipulation data, navigation data, and grasping data, and the trajectory tokenizer unifies the input.
The result: on a combined benchmark of 12 embodied tasks (manipulation, navigation, grasping, assembly, etc.), Visics's 7B model matches the task-specific SOTAs while using a single unified architecture. The model is open-sourced, including the trajectory tokenizer, the LLM, and the training pipeline.
The bigger takeaway: "Object Trajectory as Token" is a significant step toward a "general embodied foundation model." Just as text tokenization unified all NLP tasks, and pixel tokenization unified all vision tasks, trajectory tokenization may unify all embodied tasks. The "one model, many tasks" paradigm is finally arriving in embodied AI.