LingBot-VLA 2.0 scales robot pretraining to 60,000 hours

The interesting thing about the latest generation of vision-language-action models is not the model size, it is what the teams pretraining them are quietly doing to their datasets. LingBot-VLA 2.0, described in a new paper posted to Hugging Face, starts from roughly 90,000 hours of raw robot trajectory data, filters it down to 50,000 hours of what the authors call high-quality trajectories, adds another 10,000 hours of egocentric human video curated from a raw pool of about 20,000, and calls the resulting 60,000 hour mix its pretraining set. The filtering pipeline itself, which throws out episodes based on jerk, Z-scores of velocity and acceleration, video-state misalignment, and hand-pose plausibility, is presented as one of the paper's actual contributions.

Alongside the data work, the authors expand the model's action space beyond the usual dual-arm setup. LingBot-VLA 2.0 supports degrees of freedom for the heads, waists, mobile bases, and dexterous hands, all packed into a 55-dimensional canonical vector covering arm joints, end-effector pose, gripper position, hand joints, waist, head, and a mobility signal. That is the plumbing that makes cross-embodiment training work across the 20 robots in Table 1, which range from a Franka single-arm to humanoids like the Unitree G1 and Fourier GR-2. A third piece, predictive dynamics modeling, distills from a depth model called LingBot-Depth and a causal video model called DINO-Video, pushing the policy toward some future-aware temporal reasoning.

On the GM-100 benchmark under a generalist mixed-training setting, LingBot-VLA 2.0 on the Agilex Cobot Magic platform reaches a 66.2 progress score and 34.4 success rate across the nine evaluated tasks, 8.0 and 4.4 points above LingBot-VLA-1.0 and 7.1 and 2.2 points above pi_0.5. Retrieve Keychain jumps from 67.5/60.0 to a clean 100.0/100.0. On long-horizon mobile manipulation, the model edges past pi_0.5 in both in-domain and out-of-distribution runs, where OOD means the robot's start pose is perturbed by ±10 cm and a few objects are swapped for unseen categories.

The honest caveat is that on the Galaxea R1 Pro embodiment the overall success rate is only 15.6, and several tasks show a large progress-vs-success gap, meaning the model gets partway through and then fumbles the final placement. What the paper does not give you is parameter count, compute budget, or comparisons against anything other than pi_0.5 and the prior LingBot version. Take the reported wins as evidence that the data-plus-whole-body recipe helps, not that generalist robot control is solved.

For teams deploying humanoid or mobile robots this year, the useful signal is the recipe more than the leaderboard. The groups winning this round are the ones with the boring data pipelines, not the biggest models.