Robotics Briefing /
Better robots start with a clearer account of failure.
A robot that misses the last few millimeters needs a different fix from one that never starts moving. Two arms can also execute the right motions at the wrong times. Today’s papers examine these distinct failure modes, while industrial announcements show what it takes to turn autonomous movement and inspection into continuing operations.
Your quick takeaways
- Targeted human-guided corrections raised success from 12/36 to 27/36 physical trials across three precision tasks, without retraining the underlying robot policy.
- A 1,340-episode study found that the apparent winner among robot policies changed with demonstration count and training recipe. It tested one task on one inexpensive arm.
- BMW and Embotech report one million automated vehicle movements across four plants. ANYbotics has begun shipping its inspection robot for explosive atmospheres; those deliveries have a different evidence base from an established production deployment.
Papers
ARC: correct the difficult part of a mostly successful movement
Anukriti Singh and colleagues, University of Maryland and Enchanted Tools • October 27 preprint
A robot may approach an object reliably yet fail during precise alignment. Adaptive Residual Correction, or ARC, keeps the existing policy fixed and learns an additional adjustment for the moments that need it.
People first intervene during roughly eight to ten runs per task. Their intervention timing teaches a gate when to activate, while the directions of their corrections narrow the movements the robot explores. Reinforcement learning then uses task outcomes to choose the actual adjustment rather than copying each human correction.
Using Miroki’s right arm, ARC improved tray stacking from 7/12 to 12/12 successes, marker insertion from 3/12 to 7/12, and cup stacking from 2/12 to 8/12. Across the three physical tasks, success rose from 33.3% to 75%.
The result supports repairing localized weaknesses without relearning an entire behavior. However, each comparison contains only twelve trials per task. The switching gate also depends on task-specific examples and when a particular operator chooses to intervene; it is not yet a general detector of every situation needing help.
Demonstration scaling: model rankings depend on the training setup
Simon Cowling, fysical.ai • October 27 preprint
This study compared ACT, Diffusion Policy, SmolVLA and π0.5 on one physical SO-101 arm. Training sets ranged from twelve to 200 demonstrations of placing a green ball in a basket while ignoring two red balls. The researcher scored 1,340 episodes.
With twelve demonstrations, fine-tuning π0.5’s action component achieved 16/40 successes versus ACT’s 1/40. At 100 demonstrations, ACT reached 37/40 while π0.5 stayed near 70%. The study does not fully resolve the size of ACT’s advantage at larger data budgets.
Preparation mattered too. Removing an idle opening from demonstrations raised Diffusion Policy’s rate of starting to move from 4% to 90%. Starting is not completing the task, but a failure to begin can otherwise conceal the policy’s manipulation ability.
The useful lesson is that data, training choices and failure categories belong beside a model’s score. This is not a universal ranking: the study used one task, one operator and fixed training lengths, so more demonstrations also meant fewer repetitions of each. Lighting differed between policies, and small test blocks produced noisy comparisons.
Which Hand Was Late? A failed action can have two causes
Baptiste Bonin, Caro Strickland and Audrey Durand, Université Laval and Mila • October 27 preprint
Imagine one robot hand turning a key while the other pushes a door. Both gestures can be correct individually, yet fail because they do not overlap. The push may also be too weak, leaving two problems to correct.
Using a scripted two-arm task in simulation, the researchers compared what images, arm positions and force-related signals revealed. Measurements from both arms helped identify timing failures. In some cases, passive observations could not establish whether inadequate force was an additional cause.
A recovery rule first applied the correction suggested by the observable timing, then tried both timing and force corrections if necessary. Within two attempts it repaired 297 of 386 blocked trials, compared with 211 for a strategy choosing a single cause. It matched a strategy given the true cause over that two-attempt budget.
This is evidence for acknowledging incomplete diagnoses, not a demonstrated recovery system for physical robots. The results depend on a simulated mechanism, injected faults and scripted timing. Learned policies, unscheduled delays and real hardware remain to be tested.
News
BMW and Embotech report one million automated vehicle movements
October 27 announcement • Company-reported production results
BMW and Embotech report that their Automated Driving In-Plant system has moved one million vehicles across Dingolfing, Leipzig, Regensburg and MINI Plant Oxford. They describe roughly 3,500 daily movements, round-the-clock operation and availability above 99.5%.
Embotech supplies perception, prediction and planning software alongside a separate deterministic safety architecture that it says is independently certified by TÜV SÜD. The system operates in defined production and logistics areas.
The significant development is continued operation across several sites, supported by deployment, monitoring, diagnosis and recovery processes. Moving vehicles is only part of the job; keeping that flow available matters to the surrounding production line.
These are company-reported figures. Availability is not the same as an intervention-free trip rate or a collision-risk measurement, and the announcement does not provide a full definition or audit of that metric. Factory-area operation also does not establish general public-road autonomy.
ANYmal X starts shipping to customers in hazardous industries
October 27 announcement • Customer deliveries
ANYbotics says it is shipping ANYmal X to customers including Equinor and DSM-Firmenich. The legged robot is designed for routine inspections in explosive atmospheres at oil, gas and chemical facilities, following an early-access program that began in 2023.
The system combines thermal, visual, acoustic and gas sensing with the company’s Shift platform. Inspection findings are associated with particular assets and connected to plant maintenance software, helping turn a robot’s observations into work a maintenance team can act on.
The new development is customer delivery. ANYbotics’ cited figure of more than 120,000 monthly inspections describes its wider inspection system, not measured output from these newly delivered ANYmal X robots.
The announcement establishes a commercial rollout, while long-term reliability, intervention requirements and economics for the new installations remain to be demonstrated. The operational value will depend on repeatable inspection coverage and useful findings as well as the robot’s ability to traverse a facility.
Read this first
Read the demonstration-scaling study’s discussion and limitations. It shows why a robot’s score needs its training recipe, test conditions and failure breakdown beside it—and why a short run can make the same model look better or worse.
Read the paper