Robotics Briefing /
Better robot instructions start with better evidence.
A useful robot needs to know more than what movement looks right. It needs the command that produces a firm grasp, the force that reveals a stuck handle, and the operating rule that closes an otherwise clear aisle. Today’s papers examine those missing inputs—and show why providing them does not eliminate the need to check the outcome.
Your quick takeaways
- DexPort records demonstrations using the same robot hand that later performs the task. Training on commanded actions instead of measured joint positions raised average final success from 11% to 64% in three physical tasks.
- FORCA uses force measurements to revise robot programs between attempts. Its refined programs opened drawers and operated a door handle in 10/10 trials each, but development still required repeated physical testing.
- A language interface translated workplace rules into navigation policies. Even its strongest tested model accepted three incorrect policies out of 150 commands: a valid format does not guarantee the intended meaning.
Papers
DexPort: record the command that makes the grasp work
Amber Xie and colleagues • October 23 preprint
Collecting demonstrations for a many-jointed robot hand is cumbersome. DexPort lets a person carry the actual robot hand, move its wrist and select coordinated finger skills. The same hand-and-camera module then mounts on a robot arm, reducing differences between demonstration and execution.
The important distinction is between the finger position requested and the position reached. An object stops fingers before they reach their targets; that difference helps produce gripping force. Training on measured positions alone can lose the instruction to keep squeezing.
Across three physical tasks scored by final success, recorded commands raised average performance from 11% to 64% compared with a measured-state baseline. Each policy had 15 episodes per task. A fourth task, screwdriver manipulation, used holding duration and rotation speed rather than a success percentage.
The researchers collected more than 2,500 demonstrations, including apartment and office examples. Portability could broaden where training data comes from, but the device’s weight limits continuous use and rotation-heavy motions. Its assisted skills cover only part of dexterous behavior; these results do not establish unrestricted hand manipulation.
FORCA: let contact evidence change the robot program
Xiao Tan and colleagues, UIUC and Tsinghua • October 22 preprint
A camera can locate a handle without revealing exactly where it catches or how hard to pull. FORCA gives a coding agent wrist force-and-torque measurements alongside images. It writes programs that probe contact, adjust motion and maintain force while a task runs.
The model is not queried for every movement. The generated program handles live feedback, while the agent examines recorded forces and visible outcomes between attempts to revise its logic. Existing perception, grasping and motion modules provide the starting tools.
After refinement, its fixed programs succeeded in 10/10 drawer-opening trials and 10/10 door-handle trials, versus 4/10 and 1/10 for the same agent without force access. The study also tested wiping, peeling and pushing an object toward a target.
Those results measure the finished programs, not first-attempt learning. Revisions take several minutes and rely on physical trials and resets. A wrist sensor measures combined forces without identifying each finger’s contact, and longer tasks or tighter force requirements remain difficult. The useful advance is an inspectable program that changes in response to physical evidence, with development costs still to reduce.
Natural-Language to Nav2: an empty aisle can still be off limits
Jiyeong Chae and colleagues • October 23 preprint
A mobile robot’s sensors may see an open aisle while an operator knows it is reserved for loading. This interface translates ordinary instructions into explicit destinations, restricted areas and time- or event-dependent rules for Nav2, an existing robot navigation system.
A local language model produces a structured policy. Deterministic checks accept or reject it, then map adapters apply its rules without replacing the robot’s route planner or motion controller.
On 150 researcher-written English commands, the strongest tested model produced 142 correct accepted policies, three incorrect acceptances and five rejections. Accepted errors included an omitted stop and an omitted restricted zone. Structural checks caught malformed requests but could not catch every wrong interpretation.
Physical TurtleBot3 tests showed route changes and shorter travel times with operating information supplied. However, those navigation experiments used correct policies; they did not measure the full chain from arbitrary language to compliant motion. The baseline also lacked that operating information. This is a useful way to express workplace rules, with operator review and independent instruction testing still needed.
News
Eka raises $150 million for manipulation that uses vision and force
October 20 announcement • Catch-up • Company-reported capabilities
Eka Robotics announced a $150 million Series A led by Andreessen Horowitz and F-Prime. It is developing a vision-force-action model that learns through simulation and reinforcement learning, then uses visual and force information during physical manipulation.
The company reports picking varied objects outside its training distribution at 1,500 operations per hour, and demonstrations include light-bulb installation and wire insertion. These are company claims; the funding announcement does not provide a reproducible evaluation protocol or long-duration customer reliability results.
The approach matters because appearance alone cannot reveal whether a grasp is slipping or an insertion has seated correctly. Funding can support broader training and deployments, but a fast picking demonstration does not establish equal performance across different contact tasks.
The next meaningful evidence would connect speed with sustained successful work: failures, damage, interventions and recovery over extended operation. Eka says the financing will expand its model to more tasks and accelerate customer deployments; that remains the work ahead.
Read this first
Watch DexPort’s barcode-scanning comparison. A hand can reach a plausible finger position yet fail to hold the package. The side-by-side example makes it easier to understand why a demonstration must preserve the command behind the contact, not only the visible pose.
Read the paper