Bifrost AI (@bifrost_ai)

Pinned post

Early access for Manifold is now open: manifold.bifrost.ai/#contact

www.bifrost.ai Bifrost Manifold · Accelerate Robotics Research Manifold runs thousands of GPU-parallel robot policy evaluations with custom rubrics and agentic failure analysis. Insights in 30 minutes, not days.

Charles Wong Verified account @charleswongzx Aug 12

Today we are opening early access to Manifold.

An evaluation platform for robotics researchers that runs automated failure analysis on your policies.

One harness for any benchmark, simulator, or world model, with native support for @NVIDIA's Isaac Lab Arena.

Connect a policy and it runs thousands of GPU-parallel rollouts in 30 mins.

View quoted post on X

Bifrost AI @bifrost_ai 2d

@NVIDIA Isaac Lab-Arena’s randomization capabilities make it easier than ever to evaluate policy generalization across object types, spatial layouts, and physics variations.

Congrats to @RLWRLD and @NVIDIA Robotics on bringing DexBench to the community at CoRL.

It’s now live on Manifold, so you can run the accelerated version without any setup, GPU provisioning, or simulation infrastructure to manage.

Every eval is accelearted, has full 3D replays and in-depth failure analysis with no setup.

To establish an initial reference point, we've evaluated NVIDIA GR00T N1.7 on DexBench using Manifold.

Free credits to start: docs.manifold.bifrost.ai/benchmarks/dex…

docs.manifold.bifrost.ai DexBench | Manifold Evaluate robot policies on DexBench, RLWRLD's bimanual dexterous manipulation benchmark on Isaac Lab Arena, scored on frozen initial states across retail, logistics, manufacturing and hotel scenes.

NVIDIA Robotics Verified account @NVIDIARobotics 3d

A finer touch for humanoid robots. 🤖

At #CoRL2026, NVIDIA is introducing GR00T 2, an open world action model for humanoids, alongside new open industrial benchmarks from @RLWRLD_ai and @UCBerkeley built on Isaac Lab-Arena.

Explore how NVIDIA’s open world action models, scalable simulation tools and agentic workflows help researchers train, test and refine robot skills, from connecting cables to assembling small components.

Discover more ➡️ nvda.ws/4ixvdZV

View quoted post on X

Bifrost AI @bifrost_ai Nov 3

In @nebiusai's latest documentary, we joined our friends at @HavocAi_USV to see how they build autonomy across sea, air and land.

We rebuild the environments their vessels face in sim, generate data around each failure, and HAVOC retrains the same day.

0:31

Bifrost AI @bifrost_ai Aug 12

Today we are opening early access to Manifold, an evaluation platform for robotics researchers.

One harness for any benchmark, simulator, or world model, running thousands of scenarios in parallel with automated failure analysis.

Connect a policy and see every way it breaks: manifold.bifrost.ai/#contact

www.bifrost.ai Bifrost Manifold · Accelerate Robotics Research Manifold runs thousands of GPU-parallel robot policy evaluations with custom rubrics and agentic failure analysis. Insights in 30 minutes, not days.

skara ✈️ CoRL Verified account @skara_io Aug 12

For robotics researchers, evaluation can be the slowest part of their research loop. Manifold makes it the fastest.

Introducing Manifold. The platform for accelerating robotics research with failure analysis that explains itself.

A month of experiments in a week.
Simulate a thousand scenarios in 30 minutes.
Dense labels on every atomic sub-goal and action.
Agents that cluster failure modes across every rollout.
The robotics research loop that moves at the speed of compute.

One eval harness for every benchmark, simulator, or world model. Native support for @NVIDIA’s Isaac Lab Arena.

View quoted post on X

Bifrost AI @bifrost_ai Jul 24

Real world testing is non-negotiable for physical AI but you can't test everything.

Start by defining every scenario your system needs to handle = your failure space.

Map it in sim first, find the gaps and then collect the data that matters.

Bifrost AI

Latest posts 5