What Is ConRFT?

ConRFT is introduced in ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy (arXiv 2502.05450) by Yuhui Chen, Shuai Tian, Shugao Liu, Yingting Zhou, Haoran Li and Dongbin Zhao, and the project page lists RSS 2025. Read: the arXiv abstract and paper text, the project page and the repository README.

ConRFT is third-party research. Firsthand is not affiliated with its authors and makes no claim that any model was trained on Firsthand data. Firsthand captures human egocentric video, not robot data: custom collection, fast and on demand, to a written spec, with depth derived from a calibrated stereo rig, 3D hand pose at 21 joints per hand and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset under a buyer-owned commercial license. Reward labels, human interventions, teleoperation corrections and preference labels are not Firsthand deliverables.

Per the abstract, supervised fine-tuning of VLA models struggles to achieve robust performance because demonstrations are limited and inconsistent, especially in contact-rich environments. The paper adds that straightforward reinforcement learning is hard on real robots because safety and cost call for high sample efficiency and risk-aware exploration.

ConRFT has two stages that share one consistency-based training objective. In the offline stage, Cal-ConRFT trains on a small set of 20 to 30 pre-collected demonstrations, combining a behavior cloning loss with Q-learning to extract a policy and stabilize value estimates. In the online stage, HIL-ConRFT keeps the same loss structure and learns from real-world interaction, using both the demonstration buffer and a replay buffer of new transitions.

A human operator can intervene with a teleoperation tool such as a SpaceMouse, and the paper says those corrections are added to the demonstration buffer.

The action head is a consistency policy, a diffusion-model-based policy that maps random Gaussian actions to actions from the expert distribution. The authors say it helps use inconsistent, sub-optimal demonstrations and, compared with a diffusion-based action head, stays computationally lightweight for inference.

In their experiments they fine-tune Octo-small on a 7-DoF Franka Emika arm, updating only the action head.

These are the authors’ results; Firsthand does not reproduce them and ranks no models. The abstract reports an average success rate of 96.3% within 45 to 90 minutes of online fine-tuning across eight real-world manipulation tasks, a 144% improvement in success rate over prior supervised methods and 1.9x shorter episode length. Metrics are reported over 20 trials per task.

The site entries cover other ways to adapt a VLA. ConRFT pairs offline-then-online RL with a consistency policy and human interventions on a real robot. Any ranking between them is Not stated here.

In the paper, the policy sees two RGB images from a wrist and a side camera plus the robot’s proprioceptive state, and a binary classifier trained on positive and negative demonstrations gives the task reward. Firsthand states human egocentric capture with stereo-derived depth, 21-joint hand pose and one task instruction per episode, and no reward fields, intervention logs or robot proprioception.

The authors name two. A binary reward classifier can be exploited through reward hacking, and the frozen encoders and transformer backbone limit how far perception can adapt. Whether ConRFT works with human first-person video is Not stated.

When a program needs real human egocentric demonstrations that public data does not cover. Firsthand does not build or train models. Confirm the fields on a Spec Pilot, a small fast first batch; the median from signed spec to first validated batch is 48 hours where coverage is deep, three to four weeks where thin.

What is ConRFT?

A published method (arXiv 2502.05450) for reinforced fine-tuning of VLA models with a consistency policy, in an offline stage and an online stage with human interventions. It is third-party research.

What is the consistency policy in ConRFT?

Per the paper, a diffusion-model-based action head that maps random Gaussian actions to expert-like actions and is lightweight at inference.

What did the authors report?

An average success rate of 96.3% within 45 to 90 minutes of online fine-tuning on eight real-world tasks, per the abstract. These are the authors’ results; Firsthand does not reproduce them.

Is ConRFT open source?

The official GitHub repository carries an Apache License 2.0 and is built on HIL-SERL, with examples that fine-tune Octo on a Franka robot.

Does Firsthand use ConRFT, and can I request a custom dataset?

Firsthand is not affiliated with the ConRFT authors and claims no use, training or compatibility. Reward labels, interventions, teleoperation corrections and preference labels are not Firsthand deliverables. You can request a custom dataset.