Robotics Briefing /
Using the whole arm, shortening the wait, and testing what breaks.
A robot’s intelligence has to survive contact, delays and changes to its hardware. Today’s research examines all three: using an arm as more than a gripper carrier, keeping reasoning from holding up movement, and measuring how small changes to a robot’s setup affect success. Two commercial announcements show how companies want to make these capabilities easier to access.
Your quick takeaways
- WACL uses human demonstrations to teach contact with different parts of an arm. It completed 31 of 40 constrained physical trials, but planning still took tens of seconds.
- TacitVLA learns from written reasoning without making each action wait for that text. Its benchmark latency gains are promising; they are not physical deployment results.
- LIBERO-RIG finds that strong simulated task scores can deteriorate when cameras, calibration or control timing change. Hardware consistency is part of robot reliability.
Papers
WACL: the rest of the arm becomes useful
Bo-Ruei Huang, Daniel Seita and Erdem Bıyık, USC • October 13 arXiv release • Accepted at CoRL 2026
A gripper carrying a plate cannot easily open a cabinet at the same time. WACL learns how other parts of the arm can make useful contact while the gripper maintains its assigned position or orientation. Examples include pushing a cabinet while holding a plate level and turning a faucet while keeping the gripper above the sink.
The system starts with ten bare-hand human demonstrations per task. It learns contact paths relative to important points in the scene, then uses an optimizer to choose the robot link, contact location and joint movements. A vision-language model helps express the constraints the movement must respect.
Across four constrained tasks on a physical xArm 7, WACL succeeded in 31 of 40 attempts, compared with 13 of 40 for KALM. Transferring the learned components to a Franka Research 3, with the robot model changed in the optimizer, produced 27 successes in 40 constrained attempts.
This expands what an occupied arm can do, but it is not yet responsive, continuous manipulation. Planning a single candidate took 34.6 seconds on average, mostly in two hosted model calls. Execution used short, open-loop paths without ongoing correction, and the planner modeled geometry rather than contact forces or object dynamics. Slipping and pressing too lightly or too hard remain failure modes.
TacitVLA: reasoning can help without delaying every action
Jack Lu and colleagues, NVIDIA and academic collaborators • October 13 preprint
Some robot models generate written reasoning before deciding how to move. That can improve decisions, but waiting for the text adds delay. TacitVLA uses reasoning during training while preventing the action component from depending on the reasoning tokens being generated for the current step.
For tasks that need memory, the model writes a short record while the previous action sequence executes. Later decisions can use those earlier records. Tasks without that requirement can skip text generation at inference.
On LIBERO, a simulated manipulation benchmark, a model call took 147 milliseconds rather than 495 milliseconds for the explicit-reasoning comparison, measured on one H100 GPU. On the related LIBERO-Plus robustness evaluation, success was 72.9%, close to the explicit model’s 73.5%. The speed comparison concerns model computation, not a robot completing work 3.4 times faster.
The result suggests that supervision from reasoning can improve the representations used for action without forcing a pause for every explanation. Evidence comes from benchmarks and driving evaluation, including simulation, rather than physical robot trials. Performance still depends on the reasoning used for training, and writing memory can cause a wait if it takes longer than the action being executed.
LIBERO-RIG: the same task can become harder when the setup changes
Andrea Yaoyun Cui and colleagues, Luma AI • October 13 preprint
A camera shifts, a controller runs at a different rate, or calibration drifts. The requested task is unchanged, but the robot’s inputs and movements no longer match its training conditions. LIBERO-RIG isolates this problem by varying eleven aspects of a simulated robot’s setup while keeping tasks, objects and goals fixed.
Across the eight models evaluated on all four task suites, pooled success fell from 93.2% under the nominal setup to 76.7% at the middle perturbation level and 50.9% at the highest. Camera placement was especially damaging. Training for visual variation helped with camera changes but did not reliably fix calibration and timing problems that leave the image unchanged.
The practical implication is that task competence and tolerance of hardware changes need separate measurement. These percentages are not predicted failure rates for deployed robots: the study uses one simulated embodiment, changes one factor at a time and defines its own severity levels. It nevertheless exposes weaknesses that a familiar, unchanged benchmark setup can conceal.
News
Physical Intelligence opens applications for its API beta
October 13 announcement • Company-reported examples
Physical Intelligence’s API beta lets robotics teams upload data for model fine-tuning and access hosted inference. The service takes on training compute and model-serving infrastructure, while customers still supply robot data and integrate the system with their hardware.
The announcement includes Actor’s wheel loader scooping material and loading a dump truck in what the companies describe as a commercial deployment without an operator inside. That illustrates the intended reach beyond tabletop manipulation, but the announcement does not provide a sustained-operation trial table or intervention rate.
The potential benefit is easier access to capable models across different machines. Reliability still depends on the surrounding system: Physical Intelligence explicitly reports network-connectivity challenges, and most examples use cloud inference. Access is through a beta application, not an unrestricted general release.
Promise Robotics signs a homebuilding agreement with D.R. Horton
October 13 announcement • Planned production capacity
Promise Robotics announced a multi-year commercial agreement supporting production of more than 1,250 homes for D.R. Horton. Its service connects existing building designs to production planning, robot instructions, factory operations and coordination with on-site installation.
For builders, the proposition is access to automated manufacturing capacity without developing and operating a robotics factory themselves. This makes the factory service, rather than the sale of an individual robot, the commercial offering.
The 1,250-plus figure describes the agreement’s production scope, not homes already completed. Manufacturing at the planned Phoenix facility is expected to begin in early 2027. The announcement does not establish measured per-home cost savings, construction speed or quality improvements.
Read this first
Watch WACL’s demonstrations, then read its planning-time and failure analysis. Seeing an arm open a cabinet while its gripper holds a plate makes the capability clear; the slow planning and contact failures show what still separates it from everyday use.
Read the paper