THU

RDT2-FM

Tsinghua University, China · September 2025

Tsinghua University (thu-ml)

Foundation VLAOpen weights, Apache-2.0Also written RDT2 Flow Matching
Parameters
not published
No total parameter count is published for RDT2-FM.
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-FM is the continuous-action sibling of RDT2-VQ. It reuses the same Qwen2.5-VL-7B-Instruct vision-language backbone and replaces the residual-VQ token head with a flow-matching action expert built on an improved RDT architecture, emitting the same (24, 20) relative action chunk. The practical argument for it over RDT2-VQ is cost: the model card puts fine-tuning of the FM head alone at roughly 16GB VRAM, against 32GB for LoRA and 80GB for a full fine-tune on the VQ variant. Both variants are Apache-2.0 and both depend on the UMI action convention, so calibration and wrist-camera geometry, not the model, are usually what decide success on a new arm.

Architecture

Backbone
RDT2-VQ vision-language backbone (Qwen2.5-VL-7B-Instruct)
Action head
Flow-matching action expert combining an improved RDT architecture with a flow-matching objective, producing continuous actions without iterative denoising
Action chunk
24 steps per forward pass
Parameters
No total parameter count is published for RDT2-FM. It builds on the RDT2-VQ vision-language backbone (Qwen2.5-VL-7B-Instruct, 8B total per the RDT2-VQ card) and adds a flow-matching action expert whose size is not disclosed, so no honest total can be given.
Pretraining data
Same corpus as RDT2-VQ: more than 10,000 hours of UMI-gripper manipulation demonstrations across more than 100 indoor scenes.
Embodiments
Bimanual UR5e, Franka Research 3 (FR3)

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.

Published figures. These come from the model authors, not from this site.
Inference memory, reported16 GB
Fine tuning memory, reported16 GB

RTX 4090 class card is sufficient for both inference and for fine-tuning the flow-matching head alone, which is a substantially lower bar than the 80GB needed for a full RDT2-VQ fine-tune.

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

Same platform coverage as RDT2-VQ, namely zero-shot bimanual UR5e and Franka FR3 after calibration. No separate numeric real-robot table is published for the FM variant.

Fine tuning it yourself

The cheap path is to freeze the vision-language backbone and fine-tune only the flow-matching head, which the model card puts at roughly 16GB VRAM. Output is a float32 tensor of shape (T=24, D=20) of relative action deltas, 10 dimensions per arm (3 position, 6D rotation, 1 gripper width). Action statistics and normalization must match the released ones or behaviour degrades.

Where it helps, where it does not

Strengths

  • Fine-tuning only the flow-matching head fits in about 16GB VRAM, the lowest published fine-tuning bar of any model in this comparison
  • Apache-2.0 weights
  • Flow matching avoids iterative denoising, which the authors give as the reason for low-latency control
  • Shares the zero-shot cross-embodiment behaviour of RDT2-VQ

Limits

  • No published total parameter count
  • No numeric benchmark table of its own; the RDT2 paper reports results as bar figures
  • The 'low-latency' claim is qualitative, with no millisecond number and no named GPU
  • Same UMI-convention dependency as RDT2-VQ: calibration quality and camera placement decide whether it transfers
  • The model card notes that dataset and action-statistics shifts degrade behaviour

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.