RBY

LingBot-VLA 2.0

Robbyant, China · July 2026

Robbyant (Ant Group)

Foundation VLAOpen weights, Apache-2.0Also written LingBot-VLA-V2, lingbot-vla-v2-6b
Parameters
6 B
6B total.
GPU memory
11 GB
weights at bf16, computed. Fits a 16 GB card
Inference latency
130 ms
NVIDIA GeForce RTX 4090D, one inference call with 10 denoising steps
Weights
Apache-2.0
downloadable checkpoint

What it is

LingBot-VLA 2.0 is Robbyant's July 2026 successor to LingBot-VLA, described in the arXiv report From Foundation to Application: Improving VLA Models in Practice. It grows to 6B parameters on a Qwen3-VL-4B-Instruct backbone with a sparse MoE action expert, and it triples the pretraining corpus to about 60,000 hours, of which 50,000 hours are robot trajectories across 20 configurations and 10,000 hours are egocentric human video. The key engineering change is a 55-dimensional canonical state and action vector that unifies arms, end-effectors, grippers, dexterous hands, waist, head and mobile base, so a single checkpoint spans single-arm through humanoid. It is one of only two models here with both Apache-2.0 weights and a published inference latency, about 130 ms per call on an RTX 4090D with 10 denoising steps.

Architecture

Backbone
Qwen3-VL-4B-Instruct
Action head
Flow matching action expert with sparse MoE layers, plus predictive dynamics modelling using video representation and depth estimation. The published latency figure uses 10 denoising steps.
Parameters
6B total. The action expert uses sparse MoE layers with fine-grained expert segmentation and shared expert isolation. The published checkpoint is the native-depth variant. The paper does not give a separate action expert parameter count.
Pretraining data
About 60,000 hours: 50,000 hours of robot trajectories spanning 20 robot configurations plus 10,000 hours of egocentric human video.
Embodiments
single-arm, dual-arm, half-humanoid, humanoid, AgileX Cobot Magic, Galaxea R1Pro, Astribot S1, Cobot Magic-ARX X5, 20 robot configurations in pretraining

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 6 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3222 GBTraining master weights, rarely used for inference
bf16 or fp1611 GBThe usual way these checkpoints are served
int85.6 GBQuantised, expect some loss of precision on fine motions
int42.8 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 16 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference latency, reported130 ms on NVIDIA GeForce RTX 4090D, one inference call with 10 denoising steps source

No minimum GPU or VRAM figure is published. The only hardware named anywhere is the RTX 4090D used for the 130 ms latency measurement, which implies inference fits in 24 GB but the sources do not state a VRAM number.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • RoboTwin 2.0 clean conditions, 50 tasks
    93.52%
    success rateAuthors' own table. Same split: LingBot-VLA 1.0 88.56 percent, π0.5 82.74 percent.source
  • RoboTwin 2.0 randomized conditions, 50 tasks
    92.8%
    success rateAuthors' own table. Only 0.72 points below the clean split, versus a 5.98 point drop for π0.5.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • GM-100 AgileX Cobot Magic, bimanual, generalist setting
    66.2%
    progress scoreAuthors' own benchmark. Same split: π0.5 59.1, LingBot-VLA 1.0 58.2, GR00T N1.7 36.3.source
  • GM-100 AgileX Cobot Magic, bimanual, generalist setting
    34.4%
    success rateAuthors' own benchmark. Same split: π0.5 32.2 percent, LingBot-VLA 1.0 30.0 percent, GR00T N1.7 17.8 percent.source
  • GM-100 Galaxea R1Pro, bimanual, generalist setting
    15.6%
    success rateAuthors' own benchmark. Identical to LingBot-VLA 1.0 at 15.6 percent, so version 2.0 gains progress score (34.6 versus 32.7) but no success rate on this platform.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on GM-100 bimanual tasks on AgileX Cobot Magic and Galaxea R1Pro, and on long-horizon mobile manipulation with Astribot S1 (refrigerator sorting) and Cobot Magic-ARX X5 (stove cleaning), in both in-domain and out-of-domain settings. On the Astribot S1 refrigerator task in domain the authors report 77.1 progress score and 60.0 percent success.

Fine tuning it yourself

Post-training from the released base checkpoint, same workflow as version 1.0 (safetensors weights plus config and CLI yaml). A RoboTwin post-trained checkpoint, robbyant/lingbot-vla-v2-6b-robotwin, is published alongside the base model. Data must be expressed in the 55-dimensional canonical state and action format that covers arms, end-effectors, grippers, dexterous hands, waist, head and mobile base. Mirrors are available on ModelScope.

Where it helps, where it does not

Strengths

  • Apache-2.0 weights and code, one of only two openly licensed models in this comparison and the only one without a non-commercial restriction on the weights.
  • Published inference latency on consumer hardware: about 130 ms per call on an RTX 4090D with 10 denoising steps.
  • 55-dimensional canonical state and action vector unifies single-arm, dual-arm, half-humanoid and humanoid platforms including waist, head, dexterous hands and mobile base in one checkpoint.
  • About 60,000 hours of pretraining, the largest disclosed corpus among the open models here, including 10,000 hours of egocentric human video.
  • Almost no degradation under RoboTwin randomization: 92.80 versus 93.52 percent clean.

Limits

  • GM-100 success rates remain low in absolute terms: 34.4 percent on AgileX Cobot Magic and 15.6 percent on Galaxea R1Pro.
  • On Galaxea R1Pro the success rate is identical to version 1.0 at 15.6 percent, so the generation gain is not uniform across platforms.
  • The 130 ms figure covers one inference call with 10 denoising steps. No control frequency, action chunk length or end-to-end closed-loop rate is published.
  • No VRAM figure and no minimum GPU requirement is published for inference or post-training.
  • All benchmark numbers, and the GM-100 benchmark itself, come from the model authors. No third-party evaluation exists.
  • Egocentric human video makes up 10,000 of the 60,000 pretraining hours, and the report does not quantify how much of the gain comes from it.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.