How LingBot-VLA 2.0 Filters 90,000 Raw Robot Hours

Large robot datasets contain failed demonstrations, duplicated idle time, camera blur, dropped frames and state logs that no longer align with the recorded image. Training on every recorded second can teach a policy to imitate noise or associate the wrong action with the wrong visual state.

Robbyant’s paper says it began with about 90,000 hours of robotic data and retained 50,000 hours after a redesigned cleaning process. That 44% reduction is a reminder that recorded volume and usable training signal are different quantities.

The pipeline computes derivatives of action and state signals to detect trajectories with abnormal velocity, acceleration or jerk. Thresholds are set separately for each embodiment because a compact arm, mobile manipulator and humanoid do not share the same normal motion range.

Episodes are also removed when state and action signals remain almost unchanged for more than 95% of the recording. This prevents long idle segments from dominating the dataset while contributing little information about manipulation.

A trajectory can look smooth numerically and still be unusable when timestamps, camera frames and joint states are misaligned. Robbyant projects the robot model into the image using the recorded state and URDF, then uses human review to find mismatches between the projected body and the video.

Human annotators also remove blur, severe occlusion, dropped frames and misalignment across camera views. These checks matter because a VLA policy learns associations between what the camera sees and what the robot does next.

The official table covers arms with grippers, dual-arm platforms, mobile bases, half-humanoids and humanoids with dexterous hands. Named systems include Astribot S1 and Leju KUAVO 4 Pro. Other configurations contribute different arm, body, hand, head, waist and mobility signals.

Cross-embodiment training can expose the model to shared manipulation patterns while forcing it to separate hardware-specific dynamics. A grasping instruction may be semantically similar across robots, but the joint layout, reachable workspace, control frequency and end effector differ.

A generalist robot policy needs repeated evidence that links language, visual state and action across many objects and bodies. Cleaning reduces contradictory supervision. Hardware diversity reduces the risk that the policy memorizes one camera geometry or one arm kinematic chain.

Neither step guarantees generalization. A model can still fail when the lighting, object friction, calibration, payload, camera placement or task sequence moves outside the training distribution.

The official paper and repository document a concrete data-processing pipeline and enumerate 20 robot configurations. They also report evaluations on GM-100 and long-horizon mobile manipulation across two physical platforms.

This supports the claim that LingBot-VLA 2.0 was trained from a large, heterogeneous and explicitly filtered corpus. It does not reveal every source episode or independently verify the quality label assigned to each retained trajectory.

The 50,000-hour scale does not prove reliable operation on every one of the 20 configurations, nor does it establish performance in factories, homes or public spaces. Dataset coverage is an input to training, not a deployment certificate.

Independent replication would require access to the data composition, filtering thresholds, rejected-sample distribution, benchmark protocol and model checkpoints used for each reported result. The public release provides code and weights, but not the complete 50,000-hour proprietary robot corpus.

The 50,000-hour figure refers to cleaned robot trajectories. The 60,000-hour pretraining total adds 10,000 hours of filtered egocentric human-operation video. They are not competing estimates of the same collection.

Keeping those scopes separate prevents the raw 90,000-hour robot pool, the cleaned 50,000-hour robot subset and the combined 60,000-hour pretraining corpus from being merged into one misleading number.