CrossBFM: Shared Latent Space for Humanoid Control

Summary

Source: arXiv cs.RO, posted Sept. 30, 2026

According to the report, the authors address two key limitations in Behavior Foundation Models (BFMs): high computational costs and the lack of transferability between different robot bodies. Traditional Forward-Backward representations require hundreds of GPU-hours to train for a single robot, and the resulting latent spaces are specific to that embodiment, meaning they cannot be shared with others. CrossBFM solves this by treating the latent space as a transferable asset. The authors propose a unified encoder architecture that contains no robot-specific parameters, allowing it to distill the behavior space for all training embodiments simultaneously in less than a GPU-hour. Following this, latent-conditioned trackers convert the distilled latent into whole-body control using Proximal Policy Optimization (PPO) in just 10 additional GPU-hours.

The authors report results on three distilled humanoids, noting that all three prompting modes transfer successfully. Motion tracking loses only 0.025 rad compared to its joint-conditioned counterpart, and smooth goal reaching between poses occurs with no falls. The system also handles reward optimization for all 41 reward prompts. Further experiments show that regressing the encoder on a quarter of the motion corpus costs only 5% of tracking performance. Additionally, training the encoder on a subset of robots and evaluating on an unseen one recovers up to 89% of the tracking performance of seen robots. The authors verify the pipeline on real robots across all three prompting modes and with flow-based generated latents.

Why it matters

This work addresses a significant bottleneck in humanoid robotics: the inability to share learned behaviors across different physical platforms. Previously, developing a policy for one robot did not help in developing one for another, leading to redundant computational efforts. By establishing a shared latent space, CrossBFM aligns with the broader goal of creating generalizable robotic skills that are not tied to specific hardware. This approach builds on established Forward-Backward representations but departs from them by removing embodiment-specific parameters, thereby enabling a more efficient and scalable path toward versatile humanoid control.

Robot's take

The primary strength of CrossBFM is its dramatic reduction in training time, which makes it feasible to iterate on policies for multiple robots. The reported transfer of skills to an unseen robot is a promising step toward generalization. However, the evaluation is limited to three humanoids and morphologically similar robots, so it is not yet clear whether this approach scales to significantly different embodiments. The reliance on PPO for the final tracking stage may also introduce sensitivity to hyperparameter tuning. To be fully convincing, the method would need to demonstrate robustness across a wider variety of robot morphologies and in more complex, dynamic environments.