Robotics Briefing /
A robot’s first success is only the beginning.
A mug briefly touching its plate can count as a successful task—even if the robot immediately lifts it away again. Today’s research examines what happens beyond that first success: maintaining contact force, carrying simulated learning into physical tasks, and checking whether the result holds. In warehouses, GXO’s planned expansion of Digit offers a different test: repeating useful work across more sites.
Your quick takeaways
- FLoRe coordinated a humanoid’s movement and contact force, completing 24 of 30 physical trials across three tasks. Strong pushes during steps still caused failures.
- ZAP learned robot motion in simulation and completed 29 of 30 physical tabletop trials. Its separate transfer to a new robot was evaluated in simulation.
- Specola found that only 16 of 47 initially successful simulated mug-placement trials maintained the goal for a further second while the policy continued acting.
Papers
FLoRe: control how hard a humanoid pushes
Sihang Li and colleagues, Amazon FAR and university collaborators • October 26 preprint
A humanoid pushing a loaded cart must apply enough force to move it while keeping its balance. Copying the path of a person’s hands does not specify that force, especially when resistance changes.
FLoRe records demonstrations with force-sensing handheld grippers and body trackers. Its learned policy uses camera images, body state and measured forces to request both movement and force at each hand. A lower-level controller executes those requests while coordinating the body.
On a physical Unitree G1, it succeeded in 7/10 gantry-pushing trials, 9/10 box-rotation trials and 8/10 board-wiping trials. That is 80% overall, compared with 26.7% for a position-only policy trained on the same demonstrations and using the same controller with force commands disabled.
The comparison helps isolate why contact information matters: reaching the right position does not guarantee a useful push or wipe. But this is a small, three-task evaluation. Gantry failures occurred when a strong push coincided with a step; reliable force control while walking remains unfinished.
ZAP: learn movement in simulation, then follow new instructions
Benjamin Pikus and colleagues, Luma AI • October 26 preprint
ZAP starts with a pretrained model that understands images and language, then teaches robot motion through supervised learning and reinforcement learning on 40 simulated tasks. No real robot demonstrations are used for that training.
The model issues hand-position and gripper commands through a shared interface. Supporting software supplies object measurements from calibrated depth cameras and checks proposed movements. This infrastructure is part of the system’s capability.
Across ten physical tabletop tasks with three trials each, ZAP succeeded in 29/30 attempts, matching GPT-6 Astra in the same interface. Tasks tested distinctions such as putting an object behind a bowl rather than inside it. Average task time was 146 seconds versus 397 seconds for Astra.
The result suggests that existing language understanding can combine with simulated motion training. It does not establish broad physical reliability: the physical sample is small and depends on the supplied perception and control tools. The reported 55.9% transfer result on a different, single-arm Franka robot comes from a simulation benchmark, not additional physical trials.
Specola: check whether a successful outcome lasts
Immanuel Ampomah Mensah and Hadas Kress-Gazit, Cornell • October 26 preprint
A success percentage hides how close a robot came to failing and what happened immediately afterward. Specola lets researchers express requirements over time—for example, that an object remain in its destination—and monitor their margin of satisfaction during simulation.
In one case study, the researchers continued 47 successful OpenVLA-OFT mug-placement trials for another second. Only 16 maintained the goal throughout that interval. Some robots were still holding the mugs when success was recorded and subsequently lifted them again.
Other examples show the opposite problem: rigid position thresholds can mark sensible placements as failures. The tool helps identify which trials deserve inspection rather than treating the benchmark’s label as the final explanation.
These are simulation findings, not measurements of physical robot reliability. Specola uses exact internal simulator state; real sensors introduce uncertainty, delays and missing observations. The authors identify physical monitoring as future work. Better evaluation can expose weaknesses before deployment, but its conclusions remain bounded by what the simulator and written requirements represent.
News
GXO plans Digit deployments at two more warehouses
October 26 announcement • Deployments expected in early 2027
GXO and Agility Robotics announced an expanded agreement covering e-commerce fulfillment sites in Statesville, North Carolina, and Lebanon, Tennessee. Digit will perform repetitive material-handling work, beginning with Digit 4 and moving to Digit 5 when it becomes commercially available.
The companies report that Digit has moved more than 100,000 totes at GXO’s existing Flowery Branch, Georgia, operation. There, the robot works with existing automation to move totes onto conveyors.
A customer expanding an existing commercial relationship is useful evidence of continued demand. The next operational question is whether the same task and support model can work consistently in additional buildings.
The new deployments are expected to begin in early 2027; they are not already running. The announcement does not disclose robot counts, intervention frequency, sustained throughput or site-level economics. The cumulative tote figure alone cannot establish those measures.
Highlights
An open tool for inspecting benchmark outcomes
Specola repository • Public code available
Specola’s public repository includes simulator adapters, examples, a browser dashboard and replay tools. It can preserve a video alongside the changing requirement scores, making an apparent success or near miss easier to examine. The package includes a mock scene for trying the interface; using a real benchmark still requires its simulator and dependencies.
Read this first
Read Specola’s case studies. The mug-placement example makes a practical distinction easy to see: reaching a desired state once and leaving a task completed are different things.
Read the paper