RDT-170M
Tsinghua University, China · October 2024
Tsinghua University (thu-ml)
- Parameters
- 170 M
- 170M parameters in the diffusion transformer.
- GPU memory
- 0.3 GB
- weights at bf16, computed. Fits an 8 GB card
- Inference latency
- not published
- per action step
- Weights
- MIT
- downloadable checkpoint
What it is
RDT-170M is the small variant of the RDT family, with a hidden size of 1024 and a depth of 14, half of RDT-1B in both. It shares the pre-training mixture, the frozen SigLIP and T5-XXL encoders, the 64-step action chunk and the MIT license with RDT-1B. The authors present it as the VRAM-friendly option, but it appears in the paper only as the RDT(small) ablation row rather than as a separately evaluated system. Choose it when the 1.2B policy does not fit the budget, and expect to run your own evaluation because the published evidence for this size is thin.
Architecture
- Backbone
- No VLM backbone. Same frozen SigLIP-SO400M-patch14-384 vision encoder and frozen T5-v1.1-XXL text encoder as RDT-1B.
- Action head
- Diffusion transformer, same formulation as RDT-1B at half hidden size and half depth
- Action chunk
- 64 steps per forward pass
- Parameters
- 170M parameters in the diffusion transformer. The model card specifies a hidden size of 1024 and a depth of 14, which are half of the corresponding values in RDT-1B. The frozen SigLIP and T5-XXL encoders are the same as in RDT-1B and are not counted here, so the encoder memory cost does not shrink with the policy.
- Pretraining data
- Same pre-training mixture as RDT-1B: 46 datasets including RT-1, RH20T, DROID, BridgeData V2, RoboSet and an Open X-Embodiment subset.
- Embodiments
- same coverage as RDT-1B: single-arm and dual-arm manipulators, joint or end-effector control
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.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 0.6 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 0.3 GB | The usual way these checkpoints are served |
| int8 | 0.2 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 0.1 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 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.
The Hugging Face card describes RDT-170M as a more VRAM-friendly solution than RDT-1B but publishes no VRAM number. Note that the frozen T5-v1.1-XXL text encoder is unchanged, so the encoder-side memory problem described for RDT-1B persists unless language embeddings are precomputed.
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
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
No separate real-robot evaluation is published for RDT-170M. In the paper it appears as the RDT(small) row in the model-size ablation, not as a headline system.
Fine tuning it yourself
Identical fine-tuning pipeline to RDT-1B. This is the variant to pick when the 1.2B model does not fit the training budget, but the published evidence for it is an ablation row (RDT(small)) rather than a full evaluation.
Where it helps, where it does not
Strengths
- MIT licensed, same permissive terms as RDT-1B
- Roughly seven times fewer policy parameters than RDT-1B, described by the authors as the more VRAM-friendly option
- Drop-in compatible with the RDT-1B training and inference code
Limits
- No dedicated benchmark table or real-robot evaluation is published, only an ablation row, so its accuracy relative to RDT-1B is not quantified in any table this research could verify
- The frozen T5-XXL encoder is unchanged, so the main memory bottleneck is not removed by the smaller policy
- Inherits the RDT-1B limitation of not generalizing to unseen robot platforms without fine-tuning
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://huggingface.co/robotics-diffusion-transformer/rdt-170m
- https://arxiv.org/abs/2410.07864
- https://github.com/thu-ml/RoboticsDiffusionTransformer
- https://rdt-robotics.github.io/rdt-robotics/
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.