Modular Collaborative Robot Platforms

Franka FR3 entrenches as Physical AI's de facto benchmark

The Franka Research 3 arm appears in at least eight distinct arXiv Physical AI papers over the 90-day window — covering contact-rich manipulation, force-torque sensing, dexterous hand mounting, teleoperation docking, and cross-embodiment transfer — cementing its position as the default hardware reference platform for academic and applied research teams. Its combination of open software interfaces and force-control fidelity makes it the most-cited single arm in the dataset. Critically, the FR3's appearance in cross-embodiment transfer experiments (e.g., UR5-to-FR3 in [44][45]) shows it is now used as the 'target' platform researchers validate against, not merely a convenient lab fixture. This creates a powerful network effect: the more papers use FR3 as the baseline, the higher the cost for any competing arm to displace it.

UFactory's xArm series (xArm6, xArm7, and dual-arm xArm configurations) appeared in six separate research papers, spanning bimanual grocery packing [15], contact-rich πR² policy experiments [30], BPC viability on industrial arms [0], and multi-arm VLA training corpora [33]. The xArm's affordability and open SDK lower the barrier for teams to move from simulation to real hardware, and its repeated appearance in production-adjacent contexts — modified grippers, dual-arm setups — signals that researchers are treating it as a deployment-grade platform, not just a prototyping toy. This trajectory is reinforced by the xArm19's inclusion in NeoData's multi-robot pretraining corpus alongside Franka and Flexiv [33].

Why it matters · UFactory's open-ecosystem strategy is winning research mindshare faster than traditional sales cycles, creating a pathway to commercial volume as Physical AI policies trained on xArm hardware are directly deployable in SMB manufacturing.

Robotiq's 2F-85, Hand-E, and 3-Finger grippers appear across six papers as standard end-effectors, but the nature of their use is shifting: researchers are now actively augmenting them with external force sensors (Paxini sensors on Robotiq fingertips [48], wrist-mounted F/T sensors between the robot and gripper [26]), tactile arrays, and cross-morphology transfer policies [21]. Signal [21] is particularly telling — a Delto DG-3F-B gripper not in training data achieved 78% success by mapping Robotiq 3-Finger predictions, showing Robotiq's grip policy is now a transferable abstraction layer rather than a proprietary advantage. Robotiq's plug-and-play positioning is being commoditized by cross-embodiment AI.

Why it matters · Robotiq must move up the stack into sensing and software to defend margin, as pure hardware differentiation erodes when any gripper can inherit its policies via morphology mapping.

Multiple papers demonstrate that policies trained on one robot arm now transfer to another with minimal re-engineering — UR5 to FR3 [44], Franka Panda to Robotiq 3-Finger to Allegro Hand [22], and Delto gripper via Robotiq mapping [21]. AgiBot's V-Link product [18] outperforms Physical Intelligence π0.5 on LIBERO-Plus benchmarks and Xiaomi Robotics-0 on RoboTwin 2.0, signaling that the model layer is becoming the differentiator, not the arm. The NeoData corpus explicitly spans ARX X5, UR5e, Flexiv, Franka, and Piper platforms in a single pretraining run [33], formalizing multi-embodiment training as standard practice.

Why it matters · Hardware exclusivity is no longer a defensible moat — cobot OEMs that do not build or partner on AI model stacks risk becoming interchangeable commodity suppliers.

The entire 90-day capital picture for this theme is a single data point: AgiBot's $200M Series C at a $1B valuation led by RoboStrategy in early July 2026 [38]. Every subsequent week registered $0 in new funding across zero deals, even as research mention volume held steady or grew — peaking at 10 mentions the week of August 24. The velocity metric of -0.44 confirms the theme is cooling from a capital perspective despite sustained academic engagement. The stage mix is exclusively Series C [38], suggesting late-stage consolidation rather than a broad funding wave.

Why it matters · The disconnect between rising research adoption and frozen private capital suggests the next funding catalyst will be commercialization metrics — units shipped, recurring software revenue — rather than pure technical progress.