THU

RDT2-VQ

Tsinghua University, China · September 2025

Tsinghua University (thu-ml)

Foundation VLAOpen weights, Apache-2.0Also written RDT2, RDT 2
Parameters
8.3 B
The Hugging Face card lists 8B parameters.
GPU memory
16 GB
reported by the authors
Inference latency
not published
per action step
Weights
Apache-2.0
downloadable checkpoint

What it is

RDT2-VQ is the Tsinghua group's successor to RDT-1B and a complete architectural break from it. Instead of a bespoke diffusion transformer it fine-tunes Qwen2.5-VL-7B-Instruct to autoregressively emit discrete action tokens produced by residual vector quantization, predicting a 24-step relative action chunk at 30 fps (0.8 seconds) in a 20-dimensional bimanual end-effector space. The training data is not robot teleoperation but more than 10,000 hours of UMI-gripper human demonstrations from 100-plus indoor scenes, which is what enables the headline claim of zero-shot deployment on unseen arms such as bimanual UR5e and Franka FR3. It is Apache-2.0 and, rarely for this class, publishes concrete VRAM tiers (16GB inference, 32GB LoRA, 80GB full fine-tune). The catch is that everything is expressed in the UMI convention, so camera placement, calibration and gripper geometry decide whether the zero-shot claim holds on your hardware.

Architecture

Backbone
Qwen2.5-VL-7B-Instruct
Action head
Autoregressive prediction of discrete action tokens produced by Residual Vector Quantization (RVQ)
Action chunk
24 steps per forward pass
Parameters
The Hugging Face card lists 8B parameters. The paper describes the system as built on a 7B parameter VLM, so the 8B figure includes the action-token machinery on top of the Qwen2.5-VL-7B-Instruct backbone. Tensor type is BF16.
Pretraining data
More than 10,000 hours of human manipulation demonstrations collected with an enhanced, embodiment-agnostic Universal Manipulation Interface (UMI) gripper across more than 100 indoor scenes and households. The paper describes this as one of the largest open-source robotic datasets.
Embodiments
Bimanual UR5e, Franka Research 3 (FR3), UMI gripper data collection rig

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 8.3 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
fp3231 GBTraining master weights, rarely used for inference
bf16 or fp1615 GBThe usual way these checkpoints are served
int87.7 GBQuantised, expect some loss of precision on fine motions
int43.9 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 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 memory, reported16 GB
Fine tuning memory, reported32 GB

RTX 4090 class card (16GB VRAM and 32GB system RAM) is the stated inference floor. LoRA fine-tuning needs an A100 40GB class card, full fine-tuning an A100 80GB, H100 or B200. Tested on Ubuntu 24.04.

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

No results in this category are published for this model.

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Demonstrated zero-shot on Bimanual UR5e and Franka Research 3 setups without embodiment-specific fine-tuning, on unseen objects, scenes and instructions. The paper also reports dexterous, long-horizon and dynamic downstream tasks including table tennis. Numeric per-task success rates appear only as bar figures in the paper, so this research could not verify exact values.

Fine tuning it yourself

Two supported paths: LoRA fine-tuning on a 32GB-plus card, or full fine-tuning on 80GB-plus. Data must respect the model's action normalization statistics and RVQ codebook, and the model card warns that a mismatch in normalization or RVQ code version breaks behaviour. The action space is relative, so gripper geometry and camera extrinsics must be calibrated to the UMI convention.

Where it helps, where it does not

Strengths

  • Apache-2.0 weights on an 8B model, unusually permissive at this scale
  • Explicit, published VRAM tiers for inference, LoRA fine-tuning and full fine-tuning, which almost no competing VLA publishes
  • Zero-shot cross-embodiment deployment is the design goal and is demonstrated on two unseen arm platforms
  • Trained on 10,000-plus hours of UMI data, which decouples data collection from any specific robot

Limits

  • Bound to the UMI action convention: the model card states that open-world robustness depends on calibration quality, camera placement and gripper specifics
  • Input is binocular wrist-camera RGB, so a robot without the expected wrist camera geometry needs re-rigging, not just fine-tuning
  • No readable numeric success-rate tables in the paper, only bar figures, which makes head-to-head comparison against pi-0.5 or GR00T unverifiable from the primary source
  • No published millisecond latency for the base model; the speed claim applies to a distilled UltraFast variant
  • The model card explicitly requires human supervision and hardware interlocks for safety-critical use

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.