LimX COSA 0.5: Whole-Body VLA That Tidies a Room Uncut
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.
| Layer | Job | Rate (reported) |
|---|---|---|
| V³ slow system | Vision-language model reads head/wrist cameras + language → intent latent | < 5 Hz |
| V³ fast system | Flow-matching policy turns latent + live vision + robot state into short whole-body action chunks | 50 Hz |
| LimX-WBT (S0) | Single whole-body tracker converts targets to joint commands while holding balance | Control 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.
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.
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.