Tsinghua University TSAIL (thu-ml)
RDT2-VQ
8.29B Commercial use OK 8/8 verified
Facts
- Released
- 2025-09 1 Source
- Parameters
- 8.29 B 1 Source
- Architecture
- Monolithic autoregressive 0 Source
- License
- Apache-2.0 (code and model card) 2 Source
- Commercial
- Commercial use OK 0 Source
- Weights
- huggingface.co ↗ 1 Source
- Code
- github.com ↗ 1 Source
- Paper
- arxiv.org ↗ 0 Source
- Input modalities
- rgb · language 0 Source
- Action space
- 24-step × 20-dim relative action chunks (dual-arm end-effector) 0 Source
- Cross-embodiment
- 10,000+ hours of UMI human manipulation data across 100+ indoor scenes; zero-shot on dual-arm UR5e / dual-arm Franka FR3 1 Source
- Pretraining data (hours)
- 10000 h 1 Source
- Target forms
- Dual arm 0 Source
Architecture · Autoregressive VLA adapted from Qwen2.5-VL-7B-Instruct, Residual VQ action tokenizer; flow-matching version RDT2-FM also available; pretrained on human manipulation data collected with UMI handheld grippers, zero-shot cross-embodiment
Sources · 2 sources
release_date,license,arch_notes,cross_embodiment,pretrain_hours,code_url https://github.com/thu-ml/RDT2 Official page · Fetched 2026-09-05 · No snapshot stored yet README: model 2025-09, paper 2026-02; Apache-2.0; Qwen2.5-VL-7B-Instruct; Residual VQ; 10,000+ hours; Bimanual UR5e / FR3 / UMI gripper
params_b,weights_url,license https://huggingface.co/api/models/robotics-diffusion-transformer/RDT2-VQ Model hub page · Fetched 2026-09-05 · No snapshot stored yet safetensors 8,292,166,656；license: apache-2.0；createdAt 2025-09-22
Activity signals
Public counters fetched daily from GitHub, Hugging Face and ModelScope. Numbers only; no judgement. · fetched 2026-10-31
GitHub stars
807
+2 / 30 d
Commits · 90 d
0
last push 267 d ago
HF downloads · 30 d
1,735
all-time 38,471 · 23 likes
Weights last modified
2026-02-07
267 d ago
ModelScopeno same-name mirror found (searched by repo name)
Hugging Facerobotics-diffusion-transformer/RDT2-VQ
stars · last 30 d
Public benchmark scores
As printed in the source; protocols differ between papers. Full table with notes on the benchmarks page.
| Benchmark | Subset | Score | Source |
|---|---|---|---|
| Real-robot evaluation (self-reported)RDT2-UltraFast (fine-tuned RDT2-FM + one-step consistency distillation; NOT the VQ head) | cloth_folding_success | 77 | paper · Table 2, row 'Cloth Folding / Success Rate(%)', column RDT2 |
| Real-robot evaluation (self-reported)RDT2-UltraFast (fine-tuned RDT2-FM + one-step consistency distillation; NOT the VQ head) | table_bussing_progress | 0.6 score 0-1 | paper · Table 2, row 'Table Bussing / Progress Score(max=1.0)', column RDT2 |
| Real-robot evaluation (self-reported)RDT2-UltraFast (fine-tuned RDT2-FM + one-step consistency distillation; NOT the VQ head) | unzipping_success | 45 | paper · Table 2, row 'Unzipping / Success Rate(%)', column RDT2 |
Deployment recipe
No recipe yet. We publish only steps that actually ran on a rented GPU with a measured latency; this model has not been through that yet.
Record history
2026-09-05
Sources fetched
release_date,license,arch_notes,cross_embodiment,pretrain_hours,code_url · params_b,weights_url,license
Latency & cost
cost = p50 ms ÷ 1000 × rental ÷ 3600 × 1000
| Hardware | p50 ms | p95 ms | Peak VRAM | Cost per 1k inferences ¥ | samples |
|---|---|---|---|---|---|
NVIDIA RTX 4090 ¥1.68/h · GongJi Compute (2026) | — | — | — | — no latency data | No data |
NVIDIA RTX 5090 ¥2.4/h · Runyun 2.4 / GongJi 2.5 / Paratera Cloud 2.98 (2026) | — | — | — | — no latency data | No data |
NVIDIA A100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA Jetson AGX Thor no rental price | — | — | — | — no latency data | No data |
NVIDIA H100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA A800 80GB no rental price | — | — | — | — no latency data | No data |
No latency measurements for this model yet · Data comes from literature, maintainer runs, or crowd CLI uploads. None exist so far.
Compatibility
Colored only where we hold evidence. An amber dot means the vendor declares this form as a target; that is a claim, not a verified compatibility.
| Unitree G1 | Unitree H1 | AgiBot Genie G1 | AgileX Piper arm | SO-101 | Franka FR3 / Panda | Fourier GR-2 | WidowX 250 S | Google Robot（Everyday Robots） | ALOHA 2 / Cobot Magic (dual-arm) | Galaxea R1 Pro | Franka Emika Panda | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RDT2-VQ |
Legend Officially supportedCommunity reproducedFine-tunable in theoryUnsupportedUnknown Vendor-declared target form 1 / 12 cells with a definite status
Reproduction records
Did this model run on your robot? Report it with a link (log, video or repo). Records are reviewed by the maintainers; a verified “works” or “works after fine-tuning” upgrades the matrix cell to “community verified” with your link as evidence. Anonymous is fine; nothing but the fields below is stored.
No verified record yet for this model.
Clear direction, grounded judgement.