OpenWAM Opens Its Robot World-Action Stack as Training Reaches Physical Tests

01OpenWAM Opens Full Robot World-Action Model Stack as 6,400-Hour Training Reaches Real-Robot Tests

OpenWAM, a new open research stack for robot world-action models, is moving a model trained on roughly 6,400 hours of human and robot data into physical-robot evaluation. World-action models inherit knowledge from video-generation systems and use embodied experience to convert it into control signals that robots can execute.

Existing systems tightly couple their generative backbone, visual representation, architecture, information flow, inference procedure and training data. That makes it difficult to isolate which choices improve robot behavior. The OpenWAM team has instead divided the design space into controllable, composable modules, then used experiments to derive principles for building its OpenWAM-α model.

The infrastructure layer, OpenWAM-Infra, provides unified training, inference, deployment and evaluation across eight simulation benchmarks. On top of it, OpenWAM-Study ran controlled experiments around three questions: what knowledge a model should inherit, how world learning and action learning should interact, and how that combination scales.

The team concluded that upstream knowledge transfers best through a sufficiently capable generative backbone and a compact but information-rich latent space. Its experiments also indicated that world-action coordination requires dedicated capacity for actions, an explicit path for information to flow from the world model to the action system, and synchronized joint denoising during training and inference.

A further finding was that embodied pretraining principally improved generalization outside the training distribution. The researchers used one-stage co-training on egocentric footage—video captured from a person’s viewpoint—and robot data, seeking to combine broad coverage of human activity with robot-grounded actions.

OpenWAM-α was assembled from those principles and pretrained on about 6,400 hours, or 518.5 million frames, of egocentric human and robot data. The team says it remained in the top performance tier across eight simulation benchmarks spanning five robot embodiments, including single-arm, bimanual and dexterous-hand systems, and reached state-of-the-art performance on the mobile-bimanual EBench benchmark.

In physical testing, the researchers report that OpenWAM-α ranked first on the bimanual RoboDojo-Real leaderboard, with twice the success rate of pi0.5, and consistently outperformed representative world-action and vision-language-action baselines on single-arm and dexterous-hand robots. The source does not provide complete success rates, statistical intervals or independent replications for every real-robot experiment.

The release includes infrastructure code, evaluation protocols, pretrained and post-trained weights, pretrained models, and data recipes.

02AuK Unifies Speech Generation and Editing, Using Four-Step Inference for a 4.5× Speedup

AuK, a new open-source speech foundation model, brings five previously separate task families under one interface: speech generation, content editing, enhancement and separation, paralinguistic editing, and acoustic editing. Users provide natural-language instructions and audio context, allowing the same system to create speech, change what was said, isolate or improve sounds, alter features such as speaking style, and modify acoustic properties.

The team released AuK’s source code and model weights after training it on approximately 3.03 billion instruction–audio instances, representing 1.95 million hours of effective supervision. The aim is to make one model handle both generation and editing, rather than requiring a different system or control scheme for each task.

AuK processes two kinds of conditions. A multimodal large language model supplies semantic information, interpreting the requested content and instruction. A variational autoencoder, or VAE—a system that compresses media into a representation suitable for generation—supplies acoustic information. That VAE was jointly trained on speech, general audio, and music.

Generation then passes through a hybrid rectified-flow Transformer. It first uses dual-stream MMDiT blocks to process its inputs separately, followed by unified single-stream DiT blocks that combine them to produce the output.

Training began with a generation-only warm-up before moving to joint generation-and-editing pre-training. The post-training stage addressed different quality problems with separate methods: human-feedback preference optimization targeted open-ended editing, while reward-based reinforcement learning was applied to speech generation.

To lower inference cost, the researchers distilled the full model using consistency initialization and task-routed Decoupled DMD. The resulting AuK-Flash completes inference in four steps without classifier-free guidance. Under matched conditions, the team reports a 4.5× wall-clock speedup over the full model.

That gain comes with incompletely documented trade-offs. The report says the full AuK system achieved leading results on zero-shot and instruction-controlled speech generation and general instruction-guided editing, while remaining competitive on signal-level restoration. But the source does not provide the hardware, absolute latency, complete quality metrics, independent evaluations, or enough detail to determine precisely how AuK-Flash’s output quality compares with the full model.

03Uno Uses Parallel Diffusion Generation to Speed Autoregressive Models, Team Reports Up to 3× Throughput

Uno is a new language-model design intended to overcome a basic constraint of autoregressive models: they generate tokens sequentially, with each token depending on the preceding text. The research team says Uno can instead propose multiple tokens in parallel while preserving the output distribution defined by an underlying autoregressive model.

The design separates parameters into two groups. Autoregressive weights are trained with the standard next-token-prediction objective and continue to define the model’s distribution. Lightweight diffusion weights, learned through an additional distillation phase that the team calls negligible in training overhead, generate several token proposals simultaneously. Uno can either be trained from scratch or added to an existing open-weight autoregressive model.

A specialized sampler called Ψ-Spec then samples from the autoregressive distribution at a fixed context length. This differs from conventional speculative decoding, which normally uses a separately trained draft model to propose tokens for the larger model. It also differs from a pure diffusion language model: Uno retains the autoregressive weights as the authority over the output rather than accepting the quality trade-offs the researchers associate with diffusion-only generation.

The team reports that Uno delivered higher throughput than the speculative-decoding methods it tested at every evaluated batch size. Its best result was a speedup of up to three times over the base autoregressive model, including at the largest batch size supported by the test device. That figure is a maximum, not a result established for every workload.

In capability tests, the researchers say an 8-billion-parameter Uno model beat the 26-billion-parameter DiffusionGemma and the proprietary Mercury 2 across all evaluated benchmarks for agentic tool use, coding and long-context reasoning. Code and model checkpoints have been released.

The evidence remains concentrated around 8-billion-parameter systems, including a Qwen3-8B base model and the from-scratch K2-Horizon-7B-Uno checkpoint. The available source does not fully specify hardware, absolute latency, individual batch sizes or performance under KV-cache pressure. Larger models, mixture-of-experts systems, state-space models and production workloads remain untested, while community discussion also disputes the originality of the core framework and its “first” claim.