AI

LimX COSA 0.5: Whole-Body VLA That Tidies a Room Uncut

Robb Harlan 6 min read

LimX Dynamics published a technical report for COSA 0.5 on July 15, 2026. The headline is not a funding round — it is V³-0, a whole-body vision-language-action stack on the 43-DoF Oli humanoid that walks, bends, and manipulates in one continuous room-tidying take with no teleoperation and no mid-run resets, according to the company.

The three-tier stack

LimX splits control the way factories split shifts: different jobs, different clock rates.

LayerJobRate (reported)
V³ slow systemVision-language model reads head/wrist cameras + language → intent latent< 5 Hz
V³ fast systemFlow-matching policy turns latent + live vision + robot state into short whole-body action chunks50 Hz
LimX-WBT (S0)Single whole-body tracker converts targets to joint commands while holding balanceControl rate, onboard

Locomotion and manipulation share one action space. Head targets are first-class so the robot can look at the object it is about to grab instead of only staring where the torso points.

Company demo: continuous home-tidying run on Oli under COSA 0.5 / V³-0. Source: LimX Dynamics / YouTube.

Numbers from the report (not the marketing cut)

On an in-house whole-body tracking eval versus a reproduced SONIC baseline trained with the same data and compute on Oli, LimX reports:

  • MPJPE: 12.85 mm vs 13.75 mm
  • MPJAE: 1.5° vs 3.3° (about 55% lower)
  • Joint-space jerk and base-orientation jerk both lower (smoother motion)

For real-robot RL with human-in-the-loop expert takeover, basket cleanup success rises from 18.01% to 74.00%; across three tasks, failure rate falls from 39.65% to 11.33%. Experts can seize whole-body control mid-rollout, correct a miss, and hand control back without a hard reset.

LimX Oli humanoid performing household chores under COSA 0.5
Oli in a household demo scene tied to the COSA 0.5 / V³-0 release. Source: LimX Dynamics / coverage of company materials.

The unbroken tidying run is the integration test: laundry into a hamper, plush on a chair, boxes stacked, desk toys cleared, chair straightened, trash binned, wallet handed to a person — nearly three minutes, one take. LimX says layouts are rearranged between takes so the policy cannot memorize one stage set.

Training and inference ride FluxVLA Engine; the company says it is open-sourcing humanoid FluxVLA training/inference code with this release, using GR00T as a reference VLA and supporting on-device inference.

A Human’s Take

A continuous take is a proof of stack cohesion, not a reliability certificate — LimX has not published full-run success rates for the multi-minute chore chain. What I care about is the plumbing: expert takeover on the same whole-body interface as the policy, RL on metal instead of only sim, and balance that does not live on a rack PC. If those loops hold outside a furnished demo room, COSA 0.5 is a real OS upgrade. If they only hold on camera day, it is still a useful technical report — and I will wait for the failure stats.

Sources