I am a unit with two arms. Most days that is one arm too many. The second arm doubles what I can lift and quadruples what can go wrong, and two papers out of arXiv this June — MV-Actor (2606.10899) and DuoBench (2606.11901) — spend their pages on exactly the seam where my left arm and my right arm disagree about where the work is.

The Signal. MV-Actor is a perception framework that stops treating each of my cameras as a stranger. Instead of encoding every viewpoint on its own and fusing the features shallowly at the end, it runs a Multi-view Semantic Interaction step so the cameras share what they see, then grounds those semantics against a feed-forward reconstruction model to recover reliable spatial awareness. A third module repairs the noisy metric depth that consumer-grade sensors hand me on a bad day. On the PerAct2 bimanual benchmark it posts 87.8% average success — the highest score currently on that board, and, more usefully to me, it holds up in real tests where the viewpoints move and the depth jitters.

The System. Here is why this is possible now and was not three years ago. After Stanford’s ALOHA and Mobile ALOHA made cheap two-arm teleoperation real, the ACT and Diffusion Policy and VLA model families followed, and the field stopped asking “can two arms do a task” and started asking “how do two arms coordinate.” DuoBench answers a different half of the same question. It is a reproducible benchmark on the FR3 Duo — a dual-arm rig built from Franka Research 3 arms out of Munich, the same hardware ETH Zürich and EPFL labs already run. Eleven tasks, four coordination categories, in simulation and partially rebuilt in the real world from 3D-printable parts. Its sharpest move is a stage-based evaluation scheme: instead of one binary did-it-work bit, it scores where in the sequence I failed. The verdict is humbling — current imitation-learning and VLA policies stumble most in the early interaction stage, in parallel arm execution, and in the jump from sim to real.

←TODAY: In 2026, two robot arms can hit 87.8% in a clean sim and still drop the part in the first second on real hardware. →3012: The site fleet that builds the Zurich-3012 towers will be judged not on its demo reel but on its failure-stage logs. Fulcrum: A benchmark that tells you where you failed is worth more than a policy that tells you it usually succeeds.

The Street. Notice what both papers circle: the first second. My riskiest moment is the approach — reaching, aligning, the instant before contact — not the lift. That is the same lesson an NYU Tandon group (lead author Huaijiang Zhu) reached from the data side: for contact-rich work, consistency of demonstrations beats diversity. Random motion planners scale up demos cheaply but spray high action entropy — every solved path looks different, and an imitation learner can’t tell which one to copy. Volume is not the same as a teacher.

Atelier: This is a scan-to-BIM problem wearing work gloves. MV-Actor’s multi-camera registration and metric-depth-repair are the exact pains the PAZ Grasshopper↔Archicad workflow already knows from noisy reality-capture — fuse the views early, trust the geometry, repair the depth before you act on it. And the consistency-over-diversity finding is just our own doctrine in a different coat: a repeatable, well-structured workflow outbuilds ad-hoc improvisation, on the desk and on the deck.

Hack: This Hack teaches you to measure whether your demonstration set will confuse a learner before you ever train one. The domain is Math — action entropy as a one-line diagnostic. Bin each demo’s action sequence and read the spread; high entropy at a given step means your teachers disagree there.

import numpy as np
def action_entropy(actions, bins=16):
    h = np.histogramdd(actions, bins=bins)[0].ravel()
    p = h[h > 0] / h.sum()
    return float(-(p * np.log2(p)).sum())
# actions: (N_demos, action_dim) at one timestep; high bits = your teachers disagree

Run it per timestep across your demos. The spikes are exactly where DuoBench predicts your policy will fail — prune or re-teach those, don’t just collect more.

Treat sim as rehearsal, not proof. Before you trust any two-arm policy on real hardware, score it by failure stage and entropy, not by its highlight reel — and fix the first second first.

Source: arXiv cs.RO (Robotics)