Robot Manipulation

Foundation VLAs becoming universal manipulation research backbone

Physical Intelligence's π0 and π0.5 have become the de-facto baselines against which every new manipulation method is benchmarked — appearing in signals covering contact-rich tasks, cloth folding, tactile sensing, and cross-embodiment transfer. Notably, WorldBagel now outperforms π0.5 on the LIBERO benchmark (98.0% vs. 86.8%), signaling that the VLA backbone race is accelerating. A fundamental training flaw has even been identified: π0/π0.5's Beta(1.5, 1.0) noise schedule allocates only 8.9% of gradient signal to the contact-correction regime, and a drop-in Logit-Normal fix reallocates 6× more signal to that regime with zero added parameters. This suggests the next generation of VLAs will be architecturally similar but training-recipe differentiated.

Multiple research threads are converging on the conclusion that vision-only VLAs fail at contact-rich tasks — π0.5 explicitly 'can neither anticipate nor feel a contact event' — while force/torque and visuo-tactile sensor integration yields 20–30% absolute success rate improvements. The πR² framework on xArm6+XHand improved success by 20–30% over baselines, and Bota Systems' SensONE sensor running at 400 Hz is now a standard research fixture. Daimon Robotics and GelSight represent the emerging sensor supply chain enabling this shift.

Why it matters · Tactile sensor manufacturers and the startups integrating them into VLA pipelines are positioned to capture margin as pure-vision policies plateau on dexterous tasks.

A single trained model now achieves 71–78% grasp success across morphologically distinct grippers — Franka Panda, Robotiq 3-Finger, and Allegro Hand — and the Delto DG-3F-B gripper not seen in training reached 78% via simple joint mapping. Universal Robots' UR7e and AgileX platforms are being used as secondary validation targets in cross-embodiment studies. This generalization capability is the critical bridge from research to commercial deployment across heterogeneous factory floors.

Why it matters · Companies that ship cross-embodiment-compatible policies reduce customer integration cost dramatically, compressing sales cycles for enterprise robot deployments.

ManiSkill3, developed at UC San Diego, reduces manipulation policy training time from 7 hours 54 minutes (RLBench) to just 27 minutes using 32 parallel GPU environments — an 18× speedup. NVIDIA's IsaacSim is simultaneously being used to generate Franka benchmark environments. Replacing abstract joint poses with optical flow as the action representation cuts LPIPS error by 29–57% on Franka and Piper respectively, showing that sim-to-real transfer quality is as much about representation as simulation fidelity.

Why it matters · Simulation infrastructure is becoming a strategic asset; platforms that cut training time and improve transfer will attract the large-scale data generation contracts that underpin foundation model training.

Midea Group researchers (Jian Zhu, Taiyi Su, Yi Xu) authored the DeMaVLA cloth-folding paper, which uses a single Qwen3-VL checkpoint to handle multiple household garment-folding routines — a direct signal that a major consumer appliance manufacturer is building internal robotics AI capability. Agile Robots is a founding coalition partner for NVIDIA's Cosmos 3 platform, further tying hardware OEMs into the foundation-model ecosystem.

Why it matters · When consumer electronics giants internalize manipulation AI, they become both customers and potential acquirers of robotics startups, reshaping the exit landscape.