Robbyant Releases LingBot-VLA 2.0: An Open-Source 6B Vision-Language-Action (VLA) Model for Cross-Embodiment Robot Manipulation

Ant Group’s Robbyant has launched LingBot-VLA 2.0, a Vision-Language-Action (VLA) basis mannequin for robots. The launch features a technical report, an Apache-2.0 codebase, and a 6B checkpoint. The analysis group targets a widely known hole: VLA fashions typically work in labs however stumble in deployment. LingBot-VLA 2.0 advances the prior model alongside three sensible axes. These are generalization, an expanded motion house, and predictive dynamics modeling.

What is LingBot-VLA 2.0?

LingBot-VLA 2.0 is a generalist robotic coverage constructed on a vision-language spine. It converts digital camera photos and a language instruction into robotic actions. The public mannequin is lingbot-vla-v2-6b, a 6B ‘native depth’ checkpoint. It makes use of Qwen3-VL-4B-Instruct because the VLM spine. Two instructor fashions, LingBot-Depth and DINO-Video, supervise coaching by means of distillation.

One inference name takes about 130 ms on an NVIDIA GeForce RTX 4090D. That measurement makes use of 10 denoising steps. The motion knowledgeable makes use of a Mixture-of-Experts (MoE) design for scaling.

Data pipeline: 60,000 hours throughout 20 configurations

Generalization begins with knowledge. The analysis group curates roughly 60,000 hours of pre-training knowledge. This covers 50,000 hours of robotic trajectories and 10,000 hours of selfish human movies. The robotic knowledge spans 20 robotic configurations, from single-arm rigs to full humanoids. The uncooked pool is bigger: about 90,000 robotic hours and 20,000 selfish hours. A redesigned pipeline filters noisy samples right down to the high-quality set.

Filtering is express and measurable. The analysis group computes third-order jerk together with velocity and acceleration Z-scores per embodiment. Episodes with irregular smoothness or over 95% static alerts are dropped. Videos are checked towards replayed states utilizing every robotic’s URDF. Annotators take away blur, occlusion, dropped frames, and multi-view misalignment. Egocentric clips cross a VLM filter, then selfish SLAM and MANO hand-pose reconstruction.

Annotation is automated with a vision-language mannequin. Qwen3.6-27B segments every video into temporally contiguous subtasks. Each subtask will get an atomic motion from a closed vocabulary of 18 classes. That vocabulary holds 15 primitive actions plus transit, idle, and different. Across the corpus, transfer and transit dominate by frequency.

Unified motion illustration

Different robots expose completely different joints, so LingBot-VLA 2.0 unifies them. It makes use of a 55-dimensional canonical vector for each states and actions. The format is fastened throughout each embodiment within the dataset.

Each arm end-effector pose makes use of XYZ coordinates plus a rotation quaternion, giving 7 dimensions per arm. Robots that lack a physique half merely pad the corresponding dimensions. This lets one mannequin management arms, fingers, grippers, waists, heads, and cell bases.

MoE motion knowledgeable

The motion knowledgeable replaces its feed-forward community with sparse MoE layers. Each MoE layer retains one shared knowledgeable together with a number of routed consultants. Only the top-Ok routed consultants activate per token, so energetic compute stays bounded. Each knowledgeable is a SwiGLU MLP with a smaller intermediate width.

Routing follows a sigmoid-based, auxiliary-loss-free technique impressed by DeepSeek-V3. A per-expert bias corrects load imbalance with out including a load-balancing loss. Routing confidence nonetheless comes from the mannequin’s authentic, unbiased affinity scores. Under matched energetic parameters, the MoE mannequin reaches decrease coaching loss than a dense baseline. It additionally reaches decrease validation motion error on GM-100 duties.

Dual-query distillation for predictive dynamics

Real execution wants anticipation, not simply response to the present body. LingBot-VLA 2.0 appends two learnable queries to the visible and textual content tokens. Qt targets the present commentary, and Qt+T targets a future commentary. The horizon T equals the motion chunk dimension.

Two academics supervise these queries. LingBot-Depth provides express geometric cues by means of depth prediction. DINO-Video provides temporally grounded semantic priors. DINO-Video is constructed on the DINOv3 spine with block-wise causal temporal consideration and 3D-RoPE. It is skilled on 5M video clips spanning web, selfish, and robotic knowledge. On the LARYBench analysis, DINO-Video leads on three of 4 metrics.

Benchmark outcomes

Robbyant evaluates the mannequin in a generalist setting on the GM-100 (Great March 100) bimanual benchmark. A single coverage is collectively skilled on 9 duties per embodiment. Results are reported as progress rating / success price.

On long-horizon cell manipulation, the mannequin is examined underneath two settings. In-domain (ID) makes use of the coaching distribution, whereas OOD perturbs pose and objects.

Gains are largest on duties needing correct object grounding. On Agilex Retrieve keychain, success strikes from 60.0 to 100.0 versus model 1.0. Some duties nonetheless present a spot between progress and success. That hole factors to failures on the remaining exact placement or launch step.

Getting began

The repository ships set up, obtain, and deployment scripts. The instance under downloads the launched weights.

# Environment: Python 3.12, PyTorch 2.8.0, flash-attn 2.8.3
python3 scripts/download_hf_model.py --repo_id robbyant/lingbot-vla-v2-6b --local_dir lingbot-vla

Real-robot deployment runs the coverage server with compiled inference.

export QWEN3VL_PATH=path_to_Qwen3-VL-4B-Instruct
python -m deploy.lingbot_vla_v2_policy 
  --model_path path_to_posttraining_ckpt 
  --use_compile 
  --use_length 25 
  --port port

Post-training makes use of LeRobot v2.1 or v3.0 datasets. The supplied instance fine-tunes on RoboTwin 2.0 throughout 50 duties. Routing can use sequence-wise auxiliary loss with z-loss, or a loss-free setup. The config additionally exposes the Muon optimizer, with AdamW because the default.

Use circumstances with examples

The expanded motion house maps to concrete deployment situations.

• Kitchen cell manipulation: Astribot S1 kinds fruits and drinks right into a fridge. This wants base motion, door opening, and object placement collectively.
• Surface cleansing: Cobot Magic-ARX X5 wipes foam off a range with a sponge. This chains greedy, wiping, and power repositioning.
• Bimanual packing and sorting: GM-100 duties embrace egg packing, instrument packing, and block sorting.
• Dexterous-hand management: Unitree G1, Fourier GR-2, and AgiBot A2 use 12-DoF fingers, not grippers.

Interactive Dynamic Explainer