FLUX 3 Action

Open action policies currently force a choice on you: World Action Models (WAMs), which predict future actions jointly with video, lead benchmarks such as RoboLab¹. Cosmos 3 Nano² succeeds on 36.8% of tasks, while Pi0.5³, the strongest open Vision Language Action (VLA) model succeeds on 28.0%. But predicting videos along with actions makes WAMs expensive. On a B200 GPU, Cosmos 3 Nano in FP8 requires about 4.7x as much processing time per second of robot motion as Pi0.5 in BF16. When you choose a WAM you pay in latency and hardware costs; when you choose a VLA you give up on almost a quarter of successful rollouts.

Running predictions jointly over video and actions allow WAMs to transfer their world understanding - obtained through large scale video pretraining - to action prediction with limited additional action data supervision. But it also results in long sequence lengths that increase their computational costs. In addition, they often rely on guidance techniques doubling the cost of every forward pass, and run on backbones with more parameters, such as Cosmos 3 having 4.85x more parameters compared to Pi0.5. Existing solutions attempt to limit the interaction between video and actions but in doing so fundamentally change the architecture and capabilities of WAMs. Other approaches rely on parallelization of guidance but lead to increased hardware costs and even then do not reach the same speed as VLAs due to the computational footprint of their larger architectures.

FLUX 3 Action (F3A) simplifies this choice. Our single step 7B checkpoint outperforms every other open policy on RoboLab at 38.3% ± 0.38 compared to 36.8% for Cosmos 3 Nano, while featuring a 1.34x through 2.28x improvement in speed per second of robot motion compared to Pi0.5 on workstation and datacenter GPUs, despite still predicting video and actions jointly. Note that this speed advantage is in real-time factor, not latency per call: As a WAM, it can predict a longer horizon of 2.13s of motion compared to 1.0s for Pi0.5. When latency matters less, our guidance-distilled checkpoint raises the state-of-the-art success rate on RoboLab to 42.2% ± 0.36. Our base and guidance-distilled checkpoints in FP8 provide a 1.52x through 3.95x speedup over Cosmos 3 Nano in FP8 across consumer, workstation and datacenter GPUs.

Two main changes led to these results: First, a backbone less than half the size of Cosmos 3 Nano, made possible through our multimodal Self-Flow⁴ pretraining which yields strong representations. Second, distillation that removes the need for a guidance pass and reduces the number of sampling steps down to a single one, without separating video from action.

While this moves the Pareto frontier of action policies across speed and accuracy, there are still tasks that all policies, ours included, fail to solve on their own. GPT 6 Astra on the other hand can solve all tasks in the benchmark of Su et al. (2026)⁵ when using maximum reasoning efforts. But a single reasoning call adds around 35.77 seconds per second of robot motion, and each success costs $13.47 and 16 minutes on average. In contrast, F3A on its own takes 21.08 milliseconds per second of robot motion and each success costs $0.09 and less than two minutes on average when running on a H200 GPU for $3 per hour. For the most part of those two minutes, F3A is sitting idle, waiting while the robot moves. In this setting, F3A could even serve 47 other rollouts at the same time without additional costs. The main caveat being that it can’t solve every individual tasks.

A promising fix is to let a reasoner, GPT 6 Astra, delegate control to an efficient policy such as F3A or Pi0.5 and only step in when needed. In the setting of Su et al. (2026), such a hybrid policy barely improves the situation when using Pi0.5. The best hybrid costs $12.28 per success against $13.47 for pure reasoning, and takes 13.5 minutes instead of 14.

With F3A as the policy however, this delegation strategy pays off. The hybrid policy still solves 90% of all episodes, including the complex tasks no pure action policy can solve on its own, and it reduces costs to $8.77 and 8 minutes per success, making it 29% cheaper and 40% faster than the best alternative configuration. Pure reasoning with maximum effort still reaches the highest accuracy, so the trade-off between reliability and cost remains, but improved policies such as F3A change where this trade-off sits.

The idea to use predictive models of future observations for decision making has a long history¹, including early demonstrations of robot control² which combined action-conditioned video prediction with model-predictive control. UniPi instead³ generates a video of the task being performed with a text-conditioned video generator and recovers executable actions from the generated frames with an inverse dynamics model (IDM). A key argument of UniPi was that pretraining the video model on large-scale non-robotic data transfers to the robot policy and improves generalization to unseen tasks.

Subsequent work retained this use of large-scale video pretraining but coupled video and action prediction more tightly than the two-stage pipeline of video generation followed by an IDM. Joint prediction of future frames and actions within a single generative model was explored by GR-1, GR-2, WorldVLA, Cosmos Policy and LingBot-VA⁴; DreamZero introduced⁵ the term World Action Model (WAM) for this approach, which subsequent systems such as GigaWorld-Policy and the policy variants of Cosmos 3 also adopt⁶. mimic-video instead retains an inverse dynamics model but conditions it on the video model’s latent features rather than decoded frames, avoiding full video synthesis at inference and yielding what the authors call a Video-Action Model (VAM)⁷. A further line treats video prediction as an auxiliary co-training task and skips video generation at inference altogether⁸.

Previously, we explored the performance of FLUX 3 as a VAM with separate action decoder in our FLUX-mimic design study. See also mimic’s continued study on the data efficiency of flux-mimic. Here, we study the performance of FLUX 3 as a WAM with joint video-action prediction as shown in Figure 3.

Pretraining and Midtraining

Pretraining followed FLUX 3 and used a mix of image, video and audio data, with video accounting for over 95% of the training tokens. To evaluate action prediction capabilities of pretraining checkpoints, we follow a protocol similar to Cosmos 3: we add randomly initialized action heads and then finetune on DROID. The resulting checkpoint is evaluated on RoboLab.

For the results in Figure 4, we used preliminary settings with a small batch size and learning rate. As such, they do not represent the maximum performance achievable by a checkpoint. But since all runs used comparable settings, it still allows us to compare their relative capabilities for action prediction.

Note that without pretraining, performance remains below 1%. With pretraining, this jumps to 11.6% and continues to improve slowly but steadily to 12.4%. The low performance of the non-pretrained checkpoint indicates that pure DROID action training is heavily data constrained, and it shows that video-heavy multimodal pretraining can indeed overcome such limitation.

Next, we consider an action focused midtraining phase, where we continue training on a mix of our pretraining data together with newly introduced action data. We allocate 36.95% of training samples to videos with audio from our pretraining data and the remaining 63.05% to video data that comes with aligned actions. Of all training samples, 19.55% come from game recordings, 13.54% from egocentric videos with hand pose annotations, 14.03% from handheld grippers and 15.93% from teleoperation across 14 embodiments. The corresponding token ratios differ slightly due to different resolutions and durations; see Table 2.

Gaming actions are represented with 64 dimensions. The first two encode mouse movement along the x- and y-axes, the next two encode the left and right mouse buttons, and the remaining 60 encode keyboard buttons. The binary button channels are dequantized during training.

Most pose-based egocentric, handheld and teleop datasets use a 50 dimensional action space, EE50. The first 25 dimensions are used for the left or primary end-effector and the last 25 for the right or secondary end-effector. Within these 25 dimensions, the first three encode translation, and the next six encode rotation as in Zhou et al. 2019¹, but using the first two rows instead of the first two columns of the rotation matrix. The remaining 16 dimensions encode either angles of a hand pose or the state of the gripper, with 0 representing closed grippers and 1 representing open grippers. Dimensions that are unused for an embodiment are zero padded.

Note that some robot datasets rely on a separate 14 dimensional action space, split into 7 dimension for left/primary and 7 dimensions for right/secondary arms. Here, we train directly on 6 joint states and one gripper state as we do not have models of the forward dynamics of the embodiment. Across all action datasets, action rates range from 5 Hz to 30 Hz.

For the EE50 action space, the end effector pose is represented relative to an anchor frame, which is the last conditioning frame. We choose the action state closest to this anchor frame as the anchor state, and represent all other end effector poses relative to this state. Hand articulation and gripper values remain absolute. Finally, we normalize each EE50 dimension using embodiment specific z-scores as in Punamiya et al. 2025² indexed by the signed position offset from anchor state.

With the EE50 action head, we can now also evaluate RoboLab performance of midtraining checkpoints directly without additional finetuning. The results of these evaluations are shown with hatched lines in Figure 4. Initially, we observe that the EE50 performance quickly increases from 0% over 7.2% to 14.7% with modest gains for the results obtained with Joint finetuning - which improves only 2.3% over its performance in the last pretraining checkpoint. Towards later steps, we see continued improvements for both EE50, reaching 17.3%, and Joint finetuning performance, reaching 18.6%. We have not yet seen performance saturating in midtraining.

Noise-to-Signal Transition for Videos and Actions

Diffusion and flow models are learning to reverse a time-indexed forward process z_t = (1−t)x₀ + tε that mixes data x₀ with noise ε ∼ N(0,1). Depending on the scale and structure of the data, the region of t where this forward process transitions between noise and signal can be different. During training, t is sampled from a timestep distribution, and its choice can affect the performance of models significantly¹.

When jointly predicting two modalities such as videos and actions, we also have to consider the interplay between the two modalities. To get a sense of the regions of signal-to-noise transition for the two modalities, we compute the x₀ reconstruction loss separately for both. We do this for multiple checkpoints trained with various timestep distributions and then take the pointwise minimum over the checkpoints to obtain the minimum-envelope x₀ loss as an approximation for the signal-to-noise profile of videos and actions. We plot these envelopes, normalized to a maximum loss of 1, in Figure 5.

We observe an S-shaped profile with a transition around σ(t) = 3.3 or t ≃ 0.964. Assuming that model capacity is spent best in regions where the signal transition is high, a reasonable approach is to use a timestep distribution whose probability density function (pdf) is proportional to the derivative of the S-shaped profile. Depending on the function we use to approximate the S-shape, we end up with different timestep distributions. When using a sigmoid function for the approximation as shown in Figure 5, we end up with a logistic distribution in logit space, i.e. the logit-logistic distributions shown in Figure 6. Similarly, when using a Gaussian CDF, we end up with a Gaussian distribution in logit space or a logit-normal distribution.

We use these estimates as guides for choosing timestep distributions in a grid search. Besides logit-normal and logit-logistic distributions, we also include the “mode-sampler” or “Waver”² timestep distribution as it is the choice in NVIDIA et al. 2026³. We formulate the location parameter for all three candidate distributions through a zero-centered base distribution followed by an application of the shifting function, t ↦ αt / (1 + (α−1)t), with parameter α. For logit-normal and logit-logistic distributions, this results in a shift of their mode to logit(t) = log α. We choose α ∈ {24, 33, 42} corresponding to the modes of the video and action estimates in Figure 6 together with one value in-between. We use unit variance for the logit-normal distribution and consider logit-logistic distributions with scale 0.5 and 0.75. For the Waver distribution, we follow NVIDIA et al. 2026 and use a shift of α = 5 and sweep over one higher (10) and one lower (2.5) value.

Besides the timestep distribution itself, we consider one additional parameter: the scale of the action representation (not to be confused with the scale of the timestep distribution). When multiplying actions by s, but not videos, we shift the relative signal-noise transition in Figure 5 between the two. When s > 1, more noise is required to destroy the signal and the action curve gets shifted to the right. This allows us to control how different modalities are modeled in relationship to one another. Consider two modalities m₁, m₂. Assuming their signal-noise profiles are roughly matched, a model for m₁, s·m₂ with s ≫ 1 would effectively approximate a factorization like p(m₁, s·m₂) = p(m₁ | s·m₂) p(s·m₂) under a joint flow matching model, because during the transition of s·m₂ from noise to signal, m₁ would still be mostly noise but whenever we get to the transition point for m₁, s·m₂ is already mostly signal. Similarly, s ≪ 1 would achieve an approximation of p(m₁, s·m₂) = p(s·m₂ | m₁) p(m₁). To explore the effect of this, we consider scaling factors s ∈ {1, 2} for the action modality. To control for effects on the loss weighting, we reweight the loss on the action modality by 1/s².

We train the different variants and report results across multiple seeds on the simple task suite of RoboLab in Figure 7. Overall, we fail to identify a clear trend in these results, but we consistently observe high performance for logit-logistic distributions with an action scaling of s = 2 and high shifts of α = 42. Among the two scales of the logit-logistic distribution, the wider one with scale 0.75 performs best and we continue with this setting.

Split Guidance for Video and Action

Given that inference of an action chunk predicts both future video as well as future actions, we can do classifier-free guidance¹ independently on each modality. We analyze the CFG behavior by running a grid-search over {Video CFG, Action CFG} hyperparameters and evaluating success rates on the full 120 task suite.

We observe a gradient trend that prefers higher Video CFG and lower Action CFG, with Video CFG = 4.0 and Action CFG = 1.0 performing the best with a score of 42.00%. We choose this setting as our preferred inference for production setting.

EMA Analysis

We observed EMA to be critical both for performance and for reliably comparing ablation sweeps. Hence, we opted to keep track of both traditional EMA and Power EMA¹ (σ_rel = 0.05 and 0.10).

F3A provides a fast system for embodied action prediction, producing robot trajectories in as little as a single forward pass on consumer hardware. It can solve a broad range of manipulation tasks and improves performance over previous systems. But some tasks require capabilities beyond intuitive control, such as decomposing complex tasks with a long horizon or making a multi-step plan. Reasoning agents such as GPT 6 Astra provide such capability and recently showed impressive control behavior¹. However, it remains impractical to run them directly as policies, with reasoning calls on average taking around 16 seconds. Nevertheless, Su et al. (2026)² demonstrated first promising results of combining GPT 6 Astra with Pi0.5. Here, we ask whether such a combination improves the efficiency of embodied reasoning, and whether improvements in action policies - as we obtained them with F3A - translate to improvements in this efficiency.

To answer this question, we replicate the setup of Su et al. (2026): For each episode, GPT 6 Astra runs as an agent that keeps track of the episode’s history. In each turn, the action policy - F3A or pi0.5 in our experiments - makes a prediction based on the current observations and state. The joint prediction coming out of this gets converted to end-effector pose through forward kinematics. This result, along with observations and proprioceptive state, is then made available to Astra. Subsequently, it can decide between one of four options: execute a subset of the policy’s predicted actions, modify between 1-5 of the policy’s actions, propose its own actions or stop. The resulting actions are executed through inverse kinematics in RoboLab and the cycle repeats. We also include a pure variant, where Astra receives no action prediction from another policy and instead always has to make its own prediction. Note that we remained as close as possible to Su et al. (2026) and were able to reproduce their reported results. We collected timings and token counts for all Astra calls to estimate efficiency and extend their results by including F3A and running all variants across three different reasoning efforts of Astra (low, medium, xhigh).

From this data, we can give a more concrete sense of the drastic efficiency difference between pure action policies such as F3A and pure embodied reasoning agents such as GPT 6 Astra: From Table 7, we find that the most cost effective Astra variant uses xhigh efforts to achieve a cost of $13.47 per success. This variant requires an average of 16m 23s per success, whereas low reasoning efforts improve this to 14m 21s. Now, to get comparable estimates for F3A, our FP8 SD checkpoint executed a total of 147802 steps computed in 4741 policy calls for 96 successes across two seeds. Thus, on a H200 we end up with a total time of 104.8s per successful call, although most of this time is spent idle by the policy as the actions get executed. With a latency of 43.81ms on the H200, it could in theory operate 48.62 rollouts at the same time resulting in an effective time of 2.16s per success. Summarized, in a single rollout scenario with a mostly idle policy, it takes around 1.75 minutes and $0.087 per success when running on an H200 with an hourly price of $3.00/GPU-h, which makes it 8.57x more time and 155x more cost effective than the most cost effective pure Astra variant. When serving multiple environments in parallel, time per success reduces to 2.16s with a cost of $0.0018 per success - an improvement of almost 400x in time efficiency and 7843x in cost. However, it must be noted that pure Astra variants as well as hybrid variants significantly improve success rates, especially for more complex tasks, such as “Pumpkins in clutter” which none of the pure action policies can solve on their own currently.

For the hybrid policies, we see that Pi0.5 can only bring modest improvements in efficiency, reducing costs to $12.28 with medium reasoning efforts and time to 13m 33s with low reasoning efforts. F3A on the other hand improves efficiency across all reasoning levels and reduces costs to $8.77 and time to 8m 08s per success in the low setting, which correspond to an improvement of 2x and 1.54x while retaining a high success rate of 90% and solving tasks of all complexities.

One of the tasks that pure action policies struggle with is “Stack the blocks in the order from bottom to top: red, blue, green, yellow”. In the video below, we show a total of three rollouts: in the top two rollouts from F3A on its own without Astra involvement, and in the bottom a rollout from the hybrid policy of F3 + Astra (low), where the overlay indicates whether F3A or Astra is in control at any given moment. We see that F3A does have the visuomotoric skills necessary to stack the blocks - however it seems to ignore the requested ordering. In the hybrid policy at the bottom, we see that Astra intervenes for a short duration, where F3A was about to pick up the yellow block to stack it on top of the already stacked red and blue blocks. A short correction of moving the end-effector closer to the green block was sufficient to let F3A continue again on its own.