Third-party test of MolmoAct 2: color-sensitive VLA that still struggles with small-object grasping
YuXiang_IRVL · x · 2026-09-03
The IRVL team ran MolmoAct 2 on their VLA-Replica benchmark setup and shared one successful case. They put "zero-shot" in quotes since the model's training data is unknown.
Preliminary observations:
- The model is quite sensitive to object color; detailed color instructions help performance.
- It struggles with precisely grasping small objects such as blocks.
More from Embodied
- Austin users find Robotaxi 45-57% cheaper than Uber, and smoother too — whurley · 2026-09-03
- Inside China's robotics supply chain: US local content can't hit the 65% FCC bar — taochenshh · 2026-09-03
- First-person Waymo ride to Ojai: roomy cabin and a Gemini voice chat that wouldn't buy 'we're underwater' — venturetwins · 2026-09-03
- Counterfactual Debugging: causal attribution over 1M steps to localize sim2real gaps in world-model agents — MichaelD1729 · 2026-09-03
- How Google's RT-2 triggered the robotics boom: Understanding AI explains VLA models — binarybits · 2026-09-03
- Dev hooks busy bar up to Gumloop to show AI agents working live — aronkor · 2026-09-03