RBY

LingBot-VLA

Robbyant, China · January 2026

Robbyant (Ant Group)

Foundation VLAOpen weights, Apache-2.0Also written LingBot-VLA-4B, LingBot-VLA 1.0
Parameters
4 B
4B total.
GPU memory
7.5 GB
weights at bf16, computed. Fits a 12 GB card
Inference latency
not published
per action step
Weights
Apache-2.0
downloadable checkpoint

What it is

LingBot-VLA is Robbyant's first open VLA foundation model, posted to arXiv on 26 January 2026 under the title A Pragmatic VLA Foundation Model. It is a 4B parameter model that couples a Qwen2.5-VL-3B-Instruct backbone to a flow matching action expert through a Mixture of Transformers with shared self-attention, and it ships with a second checkpoint that distills depth from MoGe-2. Its practical value is that everything is Apache-2.0: weights, code and the GM-100 benchmark data. The honest reading of its results is that it beats π0.5, GR00T N1.6 and WALL-OSS on GM-100 while still only reaching 17.30 percent average success rate, so GM-100 is nowhere near solved. It was superseded by LingBot-VLA 2.0 in July 2026.

Architecture

Backbone
Qwen2.5-VL-3B-Instruct, connected to the action expert via a Mixture of Transformers with shared self-attention (inspired by BAGEL). The depth variant additionally distills MoGe-2-vitb-normal.
Action head
Flow matching action expert running in a separate transformer pathway, with proprioception fed into the action expert and multi-view images plus instruction through the VLM
Action chunk
50 steps per forward pass
Parameters
4B total. Two released checkpoints, lingbot-vla-4b and lingbot-vla-4b-depth (with depth distillation), plus RoboTwin post-trained variants. The paper does not break out the action expert size separately.
Pretraining data
About 20,000 hours of real-world teleoperated data from 9 dual-arm robot configurations.
Embodiments
AgiBot G1, AgileX Cobot Magic, Galaxea R1Pro, Leju KUAVO 4 Pro, 9 dual-arm 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 4 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
fp3215 GBTraining master weights, rarely used for inference
bf16 or fp167.5 GBThe usual way these checkpoints are served
int83.7 GBQuantised, expect some loss of precision on fine motions
int41.9 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 12 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.

No minimum GPU is stated. The repository documents Python 3.12.3, PyTorch 2.8.0 and CUDA 12.8, training tested from 8 to 256 GPUs, and a throughput of 261 samples per second on an 8-GPU setup, which is 1.5 to 2.8 times faster than existing VLA-oriented codebases.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

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, depth variant
    88.56%
    success rateSimulation. π0.5 reaches 82.74 percent on the same split in the authors' table.source
  • RoboTwin 2.0 randomized conditions, depth variant
    86.68%
    success rateSimulation. π0.5 reaches 76.76 percent, so the gap widens under randomization.source

Real world

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

  • GM-100 average over AgiBot G1, AgileX and Galaxea R1Pro, depth variant
    35.41%
    progress scoreAuthors' own benchmark. Progress score credits partial task completion, which is why it sits far above the success rate.source
  • GM-100 Galaxea R1Pro, depth variant
    20.98%
    success rateAuthors' own table. Same split: π0.5 14.10 percent, GR00T N1.6 14.29 percent, WALL-OSS 6.89 percent.source
  • GM-100 average over AgiBot G1, AgileX and Galaxea R1Pro, depth variant, 100 tasks per platform
    17.3%
    success rateAuthors' own benchmark and evaluation. Baselines in the same table: π0.5, GR00T N1.6 and WALL-OSS.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on GM-100, a real-world benchmark of 100 tasks per platform, on AgiBot G1, AgileX and Galaxea R1Pro, plus RoboTwin 2.0 in simulation. The depth variant averages 17.30 percent success rate and 35.41 percent progress score across the three real platforms.

Fine tuning it yourself

Post-training is supported from the released base checkpoints. The model path must contain safetensors weights, config.json and lingbotvla_cli.yaml. GM-100 evaluation uses 130 post-training episodes per task across 100 tasks per platform, which is the scale of demonstration data the authors themselves use for adaptation. Post-trained RoboTwin checkpoints are published alongside the base models.

Where it helps, where it does not

Strengths

  • Fully Apache-2.0: weights, code and the GM-100 benchmark dataset are all released, so results are reproducible.
  • Trained on about 20,000 hours of real dual-arm teleoperation, one of the larger disclosed real-robot corpora.
  • Optional depth distillation variant gives a measurable gain, for example 4.28 percent average success rate and 7.76 percent progress score over π0.5 across three embodiments.
  • Training throughput of 261 samples per second on 8 GPUs, 1.5 to 2.8 times faster than comparable VLA codebases, which lowers the cost of post-training.
  • Degrades far less than π0.5 under RoboTwin randomization (86.68 versus 76.76 percent).

Limits

  • Absolute GM-100 success rates are low, 17.30 percent on average, so real-world generalist performance is still poor in absolute terms.
  • Pretraining covers dual-arm configurations only. The authors name single-arm and mobile robot data as future work.
  • No inference latency, VRAM figure or minimum GPU requirement is published for version 1.0.
  • No control frequency is stated anywhere.
  • Superseded in July 2026 by LingBot-VLA 2.0, which the authors report as better on every shared benchmark (RoboTwin clean 93.52 versus 88.56 percent).
  • All benchmark numbers, including the GM-100 benchmark itself, come from the model authors.

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.