Pretraining
Train world action models with our researchers using petascale multimodal interaction data — video, proprioception, force, and language.
World action models · Checkpoints · Evals
Exploring the
Frontier
Our researchers explore the frontier of multi-modal models, how to develop them, and how to deploy them for custom needs.
Work with our research lab
Work directly with our researchers on the model capabilities your team needs. Bring us a focused research problem, or work with us from data collection through deployment.
Data collection
Develop the datasets your models need with synthetic interaction data from Hedra inference or targeted real-world captures.
Synthetic · Real-world · Datasets
Post-training
Fine-tune and evaluate foundation VLAs and open model families for your embodiment and tasks.
Adapters · Evals · Rollouts
Deployment
Bring the models we develop together into production through Hedra inference, on edge hardware, clusters, or hybrid fleets.
Edge · Cluster · Hybrid
Keep ownership of your models
Own the artifacts
The data, checkpoints, adapters, evals, and deployment manifests from our work together remain yours and exportable.
Open foundations
Build on open foundations such as Gemma, Qwen, Kimi, or robotics architectures we help adapt. Keep the ability to inspect, modify, and move your models.
Choose your runtime
Run the models we develop together on Hedra clusters, your own edge hardware, or both. Hedra can operate the runtime while you retain ownership of the trained artifacts.
Use cases
Bimanual manipulation
Assembly-cell coordination for fixture loading, tool use, and dexterous handling. Contact-rich policies react at control rate across both end-effectors.
Mobility
Warehouse aisle, yard, or factory-floor mobility for legged and wheeled platforms — traverse, recover, and keep operating off the demo mat.
Perception
Line-side camera rigs, depots, and lab cells: fuse multi-camera vision, depth, and proprioception into the world state the policy acts on.
Recently from the lab
WMR leads to big gains.
In our MiniMax H3 tests, WMR, post training and other optimizations reduced generation time on the same hardware.
Time to generate one 5-second 768p clip
MiniMax H3 · One AMD MI355X
WMR Turbo · auto optimization
2.6×
WMR · same output
1.2×
SGLang
1.0×
WMR Turbo: 62% less time per clip.
2.6×
WMR Turbo · auto optimization
1.2×
WMR · same output
1.0×
SGLang
62% less time per clip
0
25
50
75
100% of SGLang
MiniMax H3 · 5-second 768p clip · One AMD MI355X
Internal benchmarks, Aug–Sep 2026. Both speedups are relative to SGLang.
Read the research
Selected work by our researchers on generative models, embodied AI, and physical dynamics.
