Behavior Foundation Models (BFMs) give humanoids a promptable policy over a latent behavior space, so that a single vector can specify a motion to imitate, a pose to reach, or a reward to maximize. Forward–Backward representations produce such spaces, but at the cost of hundreds of GPU-hours for a single robot — and a second robot repeating the procedure arrives at a second space unrelated to the first. The result is embodiment-specific latents that neither unify nor transfer under the same prompts.
We address this problem with CrossBFM, treating the latent space itself as the transferable asset for various embodiments. As retargeting already puts every robot in frame-level correspondence, transferring it becomes supervised regression, with neither simulator nor training required on the target side. We propose a unified encoder architecture for distilling behavior space from cross-embodiment correspondence, which allows a single encoder with no robot-specific parameters to address all training embodiments simultaneously in only under a GPU hour. Following this encoder, latent-conditioned trackers turn the distilled latent into whole-body control in a conventional PPO training manner in just 10 more GPU hours.
On three humanoids, all three prompting modes transfer: motion tracking with a latent-conditioned policy losing only 0.025 rad to its joint-conditioned counterpart, smooth goal reaching between poses with no falls, and reward optimization for all 41 reward prompts. Regressing the encoder on a quarter of the motion corpus costs only 5% of tracking performance, and 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, demonstrating cross-embodiment generalization to morphologically similar robots. We also verify the pipeline on real robots across all three prompting modes and with flow-based generated latents.
A BFM trained on the Unitree G1 is a structured, value-functional space that one vector drawn from it can represent a motion, a goal or a reward for the latent-condition trackers. Rather than retrain a space for every new robot which takes hundreds of GPU-hrs and ending up in totally unrelated spaces, we freeze the G1’s space and teach other humanoids to use it.
Stage 1 leverages the cross-embodiment correspondences from retargeted data on the same timeline to train a unified encoder. This encoder converts robot X's proprioception into the G1's latent space, so that the same latent can prompt every robot to perform the same behavior.
Stage 2 trains a latent-conditioned policy for each robot. This policy observes the shared latent and its own proprioception, instead of joint commands. A flow-based model trained to generate latents from this distribution can then prompt any robot to perform a behavior such as “walk”, “run”, “dance”.
Frame t of a retargeted clip is the same pose of the same behavior on every robots. So the G1’s latent for that frame is a ground-truth label for every other robot, and transfer is a cosine regression, requiring no simulator, RL or discriminator for the new robot. Training this encoder is super fast (<1hr) compared to unsupervised RL.
Different humanoids may vary in both joint and link configuration. Our encoder use a robot-independent canonical input with 33 canonical joint indices, 8 key-body indices and a root block &mdash. Link and joint mapping is processed by their functionality.
Each robot’s policy observes behavior intent [z | proprio] rather than joint-space trajectory. Because z lives in the G1’s coordinates, all three BFM-Zero prompting modes can transfer to new robots. Moreover, as the latent space is shared, a flow-based generator trained on the G1 can prompt any robot to perform a behavior such as “walk”, “run”, “dance”.
ẑt = proj(EX(oXt−T+1:t)), proj(v) = √d v / max(‖v‖, ε)
Latent projection. The encoder takes a window of T proprioceptive frames from one robot and returns a latent for every frame in it, rescaled to the sphere of radius √d.
ℒalign = 1N ∑t (1 − cos(ẑt, z⋆t))
Encoder training objective. A frame-wise cosine regression onto the frozen source latent z⋆t = proj(BS(oG1t)), averaged over the N frames. The cosine distance also serves as a proximity measure for cross-embodiment representation on the pretrained latent space.
We observe that Pre-LayerNorm helps prevent mode collapse. We also make the projection part of the training graph so that the encoder can learn to output latents that are already on the sphere, rather than relying on a post-processing step.
We distill from a frozen G1 BFM onto three humanoids using LAFAN dataset. These robots differ from G1 in DoF count, body count, joint topology and joint order.
Each latent-conditioned policy is compared with a joint-conditioned one (TWIST2) trained on the same data, environment, reward and PPO budget. The joint policy is the upper bound as it has access to the exact joint angles at every step, while ours gets a latent behavior and have to infer the pose from it.
The gap stays within 0.025 rad on all three robots (+0.025 on M3, +0.006 on T1, +0.003 on N1), with no falls. For goal reaching, a sudden jump in joint targets makes the joint policy spike its torques and fall, while interpolating in latent space gives smooth pose-to-pose transitions for at most 0.032 rad over tracking.
Two of the three prompting modes immediately transfer after encoder training. Tracking is EX on a retargeted reference while goal reaching is EX on a predefined window ending at the goal pose.
Reward optimization does not transfer directly, because the closed form that converts a reward into a latent requires two components from the frozen source BFM:
zrew = proj( ∑i ωi ri 1BS(si) ), si ∼ 2ρ
Directly substituted with the trained encoder EX, which was regressed onto proj(BS(·)) frame by frame. BS → EX
The visited states collected during RL training are not available on a target robot that never ran unsupervised RL. We approximate ρ using the robot’s own retargeted dataset, so the reward-weighted projection selects states that are feasible for it. ρ → cached retargets
Reward optimization graph
We score all 41 reward prompts against using the G1’s oracle latent optimized on G1 buffers.
| Robot | Computed on G1 | Computed on target (ours) | Gain |
|---|---|---|---|
| M3 | 0.145 | 0.508 | 3.50× |
| T1 | 0.228 | 0.261 | 1.14× |
| N1 | 0.177 | 0.523 | 2.95× |
| Mean | 0.183 | 0.431 | 2.35× |
Normalized return in [0, 1], averaged over the 41 source tasks. Higher is better.
Computing zrew on the source robot is 3.5× and 3.0× worse on M3 and N1. The reason is that the reward-weighted projection selects which states are worth visiting, which depends on which states the executing body can actually reach. Thus a latent optimal for the G1’s reachable set is not optimal for other robots.
Does the encoder work on clips it never saw? We evaluate how well the unified encoder infers latent behavior across the training and validation motions by passing to the policy either the G1’s oracle latent z⋆ or the encoder’s inference.
| Motions | M3 | T1 | N1 | |||
|---|---|---|---|---|---|---|
| oracle | pred. | oracle | pred. | oracle | pred. | |
| Training (30) | 8.14 | 8.02 | 5.70 | 5.43 | 5.58 | 5.57 |
| Validation (10) | 6.40 | 7.07 | 5.26 | 5.60 | 5.46 | 5.60 |
Closed-loop key-body error (cm, lower is better). Every episode runs full length with no falls.
On the encoder’s training clips the predicted latent is marginally better than the oracle. On validation motions it costs only 1.4–6.7 mm of key-body error. In terms of geometrical alignment, distilled trajectories overlay the source trajectories almost identically, behaviors form clear clusters, and the validation clips (dashed) fall inside the cluster of their own behavior rather than off-manifold or in the gaps between clusters. So the tracker trained on that region adapts to them well even with some cosine dissimilarity.
How much data does it take to distill the space? We shrink the encoder’s corpus by sampling 1.5–3 s segments of every motion, so each behavior stays represented, and feed its latents to a tracker trained on the full data.
A quarter of the corpus costs only 5% (0.213 → 0.224 rad); the curve breaks only below that, at 0.271 for 10% and 0.300 for 5%. In the latent map above, 25% still overlays the source with behaviors separated, while at 10% the classes blur. We conlude that what matters is behavioral coverage, not raw frame count.
We report the generalizability of the unified encoder to unseen robots by training it on two robots and evaluating on the third, with fixed pretrained trackers. Signal gap closed is (random − pred) / (random − oracle), representing how meaningful the latent signal is to the tracker, with the upper bound being the seen robot and the lower bound pure noise.
| Unseen robot | Cosine (unseen) | Oracle MAE | Predicted MAE | Random MAE | Signal gap closed |
|---|---|---|---|---|---|
| M3 | 0.625 | 0.165 | 0.183 | 0.337 | 89.3% |
| N1 | 0.586 | 0.137 | 0.224 | 0.332 | 55.5% |
| T1 | 0.253 | 0.168 | 0.289 | 0.306 | 12.8% |
Closed-loop joint MAE in radians. With all three robots in training, cosine is 0.83–0.87.
Transfer follows morphological proximity. M3 and N1 are a close pair, and each recovers most of the gap when the other was seen while T1 has no close relative and recovers little. Cosine alone undersells this insight. M3 at 0.625 looks broken yet tracks at 89% of the oracle because the policy only needs the latent to land in the right behavior region. We hypothesize that with many embodiments and every new robot has a close relative in the training set, transfer can be expected to generalize well in a zero-shot manner.
@inproceedings{crossbfm2026,
title = {CrossBFM: Distilling a Shared Latent Behavior Space
Across Humanoid Embodiments},
author = {Do, Tan-Dzung and Phuong, Tuan Dat and Bohlinger, Nico and
Trinh, Cuc T. and Ju, Siwei and Ngo, Vien Anh and
Peters, Jan and Wang, Xinchao and Le, An Thai},
booktitle = {Under Review},
year = {2026}
}