GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

LingBot-VLA 2.0 Trains One Policy Across 20 Robot Configurations — 90K Hours Filtered to 50K

The persistent problem in robot foundation models is generalisation across bodies. Most Vision-Language-Action models learn policies tied to a specific hardware configuration — swap the arm, add a mobile base, or change the gripper, and the policy breaks. LingBot-VLA 2.0 from RobbyAnt Brain is built against that constraint.

One Policy, 20 Configurations

LingBot-VLA 2.0 uses a single 55-dimensional action format that covers the full kinematic range of modern humanoid and mobile manipulation platforms: arms, grippers, dexterous hands, head, waist, and mobile base. The same model handles 20 distinct robot configurations without task-specific retraining.

The architecture adds a sparse Mixture of Experts module at the action token level. Each token uses a small subset of specialised expert networks alongside a shared expert that preserves common manipulation skills across body types. A secondary training signal asks the model to predict depth and video features — current and future — forcing it to track object geometry and likely scene evolution during the next action chunk. That makes it aware of physics, not just current state.

Dataset: 90K Hours to 50K

Raw robot data is noisy. LingBot-VLA 2.0 filters 90,000 raw robot hours through a multi-criteria quality pipeline: jerky motion, broken sensor signals, camera-body mismatch, image blur, dropped frames, and extended static periods all trigger removal. 50,000 hours survive the filter — a 44% rejection rate.

Human video is also incorporated as a training signal, but only the useful portion. Footage is screened for hand-object interaction quality, then camera motion and hand pose are reconstructed into action-compatible data. This augments robot hours with human dexterity demonstrations without inheriting the noise of unconstrained video.

Benchmark Results

On Agilex GM-100, LingBot-VLA 2.0 posts:

  • Progress rate: 66.2% (pi0.5: 59.1%)
  • Success rate: 34.4% (pi0.5: 32.2%)

The success-rate gap versus pi0.5 is narrow — 2.2 points — but LingBot-VLA 2.0 is also generalising across 20 hardware configurations simultaneously, a constraint pi0.5 does not operate under in the same form. On both long-horizon mobile tasks in the GM-100 suite, LingBot-VLA 2.0 leads pi0.5 in-domain and out-of-distribution.

Several individual tasks within GM-100 still favour other models. LingBot-VLA 2.0 does not claim universal superiority — its claim is breadth: one model that handles diverse robot bodies without specialisation.

The Structural Bet

The multi-configuration approach is a dataset and architecture bet. If robot hardware continues diversifying — more mobile bases, dexterous hands, varying arm geometries — a policy that degrades per-swap is a liability. LingBot-VLA 2.0 is betting that a model trained on diverse body data from the start is worth the 2.2-point gap against a specialised rival. That bet becomes more interesting as the robot hardware market fragments further.