From Foundation to Application: Improving VLA Models in Practice (LingBot-VLA 2.0)

TL;DR

LingBot-VLA 2.0 is the second-generation π-style VLA from Robbyant (Ant Group), advancing the LingBot-VLA: A Pragmatic VLA Foundation Model recipe along three axes: (1) a 3× data scale-up to ~60,000 hours (50K robot trajectories across 20 configurations + 10K egocentric human video hours), (2) an expanded whole-body action space covering heads, waists, mobile bases, and dexterous hands beyond the v1 dual-arm-only interface, and (3) an explicit predictive dynamics proxy task where the model forecasts future frames using a video representation model for semantic priors and a depth estimation model for geometric cues. The paper reports GM-100 benchmark gains in a generalist setting and demonstrates strong cross-embodiment long-horizon mobile manipulation, positioning itself as the practical / deployment-focused counterpart to the causal-AR unification thread (Causal World Modeling for Robot Control (LingBot-VA)) in the same LingBot family.

Key claims

Pretraining data is scaled to ~60,000 hours: 50,000 hours of robot trajectories across 20 robot configurations plus 10,000 hours of egocentric human videos, with a revamped data processing pipeline [Abstract].
The action space is expanded beyond dual-arm hardware to include degrees of freedom for heads, waists, mobile bases, and dexterous hands, enabling more complex practical scenarios [Abstract].
Future frame prediction is formulated as a proxy task for temporal reasoning, with two conditioning signals: a video representation model supplies semantic priors and a depth estimation model supplies geometric cues [Abstract].
Evaluations on the GM-100 benchmark in a generalist setting validate the beneficial impact of the three proposed modifications [Abstract].
The expanded whole-body-DoF pretraining data enables strong cross-embodiment long-horizon mobile manipulation across the two robotic platforms in the paper’s evaluation [Abstract].

Method

LingBot-VLA 2.0 keeps the π-style VLA backbone + action-expert architecture of v1 (Qwen-VL-family backbone + action head over action chunks) but rebuilds the training corpus and adds two new supervisory signals. The revamped data processing pipeline curates 60K hours of pretraining data — the robot side aggregates 50K hours across 20 heterogeneous configurations (mobile bases + dual arms + dexterous hands + head/waist DoFs), and the human side contributes 10K hours of egocentric video that the pipeline converts into training-compatible signals despite not sharing the robot’s action space. The action head is widened to cover the additional DoFs (head/waist/mobile base/dexterous hand joints) so a single model can output actions for whole-body mobile manipulation rather than just parallel-gripper wrist commands. On top of behaviour-cloning-style action supervision, the training adds a future-frame prediction proxy task: the model predicts what the scene will look like N steps ahead, supervised by (a) a video representation model’s embedding as a semantic prior and (b) a depth estimator’s output as a geometric prior — a two-headed dynamics-prediction signal rather than raw pixel loss or a full generative video decoder. Full architectural specifics (depth model, chunk sizes, per-DoF loss weighting, whether the video representation is a frozen encoder or the base VLM itself) live in the paper body beyond the abstract’s 30KB extraction window.

Results

The paper reports GM-100 benchmark gains in a generalist setting demonstrating the beneficial impact of the three proposed modifications, and strong cross-embodiment long-horizon mobile manipulation capability across the two robotic platforms enabled by the expanded pretraining data covering whole-body DoFs [Abstract]. Quantitative per-task success rates, per-configuration ablations, and comparisons to π₀.₅ / GR00T N1.6 / v1 baselines are in the paper’s experiments section beyond the extracted abstract; the v1 GM-100 numbers to compare against are 17.30% SR / 35.41% PS average across three platforms (LingBot-VLA: A Pragmatic VLA Foundation Model GM-100 table).

Why it’s interesting

LingBot-VLA 2.0 lands at the intersection of three recipes the wiki has been tracking simultaneously. First, on the action-pretraining scale axis, it triples v1’s 20K-hour robot corpus to 50K hours across 20 configs — the closest filed answer to whether π₀.₅ and RECAP (π*0.6: a VLA That Learns From Experience (RECAP)) can be matched by a fully-open lab, and roughly 2.4× larger on the robot side alone than EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data‘s 20,854-hour egocentric corpus. Second, on the human-video-as-pretraining-substrate axis it takes the “keep both channels” position by mixing 10K hours of egocentric video directly into the same pretraining rather than as a separate stage, complementing ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining‘s reliability-aware joint training and Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s drop-the-noisy-channels reduction. Third, on the world-model-as-auxiliary-signal axis it adds a two-headed (semantic + depth) future-prediction proxy — a lighter-weight version of the v1 depth-distillation from MoGe-2 (LingBot-VLA: A Pragmatic VLA Foundation Model), and structurally similar to the “video prediction as auxiliary loss” pattern surveyed by Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models — but explicitly kept as a proxy task rather than a full generative world model in the loop (Causal World Modeling for Robot Control (LingBot-VA)‘s unified-AR alternative). The paper thus reads as the deployment-focused / pragmatic sibling to LingBot-VA’s architectural unification bet, and pairs directly with LingBot-World 2.0 / LingBot-World-Infinity — Infinite Worlds with Versatile Interactions (the same lab’s whole-scene interactive world model) as the two July-2026 releases exposing whole-body manipulation from an open lab.