Robotics Briefing /
Less waiting, longer range and robot programs that learn from a demonstration.
Useful robots need more than a successful move. They need time to react, energy to keep working and a way to adapt a task to a different scene. This weekend’s research selections examine those constraints. A spacecraft mission also shows how a hardware fault can exhaust the resources needed to finish the job.
Your quick takeaways
- RAPID turns one human demonstration into a robot program, then tests and revises it in simulation. Physical success ranged from 40% to 90% across eight tasks, with ten test scenes per task.
- Rolling-WAM cut steady-state replanning time from 978 to 215 milliseconds in a controlled A100 GPU test. Its separate physical evaluation covered three manipulation tasks.
- A new Nature paper examines RAIBO2’s 2024 marathon on one battery charge. Katalyst’s LINK mission, meanwhile, ended without capturing or boosting its target satellite.
Papers
RAPID: turning a demonstration into a reusable robot program
Yuyao Liu and colleagues, MIT, NUS, Penn and NVIDIA • Submitted September 24; listed September 25 • Weekend catch-up; preprint with physical tests
A box can be too awkward to grasp directly. A person might push it toward an edge, flip it and then pick it up. RAPID uses a human demonstration and a language instruction to build a program for that kind of sequence.
The system infers the desired outcome, reconstructs a simulation and develops the movements needed to reach it. A coding agent tests and revises the program across generated scene variations. Movements depend on object relationships, helping the same strategy handle changes in position or shape.
On a physical Franka arm, success ranged from 40% to 90% across eight tasks, with ten test scenes per task. The comparison coding-agent system ranged from 0% to 30%. In a separate simulation evaluation with fifty novel scenes per task, RAPID averaged 75.9% success.
The result suggests demonstrations could supply both a task description and a way to test robot software. Program construction takes tens of minutes per task, and transfer depends on reconstructed geometry and estimated physics. Eight tasks on one arm leave broader reliability unresolved; the project lists code as coming soon.
Rolling-WAM: reusing predictions to update movements faster
Yinghua Zhou, Junjie Ye and colleagues, USC and collaborators • Submitted September 24; listed September 25 • Weekend catch-up; preprint with physical tests
Some robot models predict future video alongside movements. That helps them anticipate what comes next, but generating a fresh prediction can delay each action update. Rolling-WAM carries partially refined predictions forward and updates them as new camera images arrive.
The model finishes its next movement first while continuing to refine more distant predictions. In a controlled RoboTwin setup on one NVIDIA A100 GPU, updates took 215 milliseconds versus 978 for Joint-WAM. That roughly 4.5-fold speedup excludes startup and warm-up.
A separate Unitree G1 experiment tested placing a toy, stacking plates and pouring beads. Each task used fifty training demonstrations and twenty evaluation trials per method. Rolling-WAM averaged 85% success, compared with 78.3% for Joint-WAM.
Faster updates could reduce pauses and help robots respond during manipulation. The timing result does not establish the same speed on smaller onboard computers. Retained predictions can also fall behind rapid scene changes, and the physical tests cover only three tasks. The repository says code and checkpoints are still in preparation.
RAIBO2: the engineering behind a marathon on one charge
Choongin Lee, Donghoon Youm and colleagues, KAIST • Nature, September 23 • Catch-up; peer-reviewed study of a 2024 demonstration
RAIBO2 completed 42.195 kilometers in 4 hours, 19 minutes and 52 seconds on one battery charge. The marathon happened in 2024; this week’s development is the Nature paper explaining the robot’s efficiency.
The team addressed losses across the whole machine through lightweight mechanics, lower-resistance motor electronics and a movement policy designed to waste less energy. The publisher reports 1,280 watt-hours consumed during the run.
Longer operating range could make legged robots more useful for outdoor inspection and search. A marathon result does not establish performance while carrying rescue equipment or crossing disaster debris.
Access was limited to the paper’s abstract and public supporting sections, plus the publisher’s summary. The linked repository contains locomotion-training and reward-evaluation code.
News
LINK ends its satellite-servicing mission without the planned capture
Katalyst Space • September 25 • Company mission report
Katalyst says its LINK spacecraft reentered the atmosphere after 85 days in orbit. The mission aimed to capture NASA’s Swift Observatory and raise its orbit, but never reached that stage.
During commissioning, an electrical fault left two of three reaction wheels unavailable. Engineers regained limited orientation control using the remaining wheel and thrusters. The extra fuel use left too little propellant to safely attempt the capture and orbit boost.
The company reports tests of its three robotic arms and grippers, plus an approach within 12 to 15 kilometers of Swift. Those checks provided flight experience without demonstrating satellite capture.
For robotics, the lesson is about the whole system: recovering control can still consume the resources needed to complete a task. Katalyst’s account documents that tradeoff, but it remains the operator’s report and does not establish a proven servicing capability.
Highlights
An IROS reading list on combining robot skills
Compositional and Modular Learning workshop • Scheduled September 27 • Upcoming event and public paper list
An IROS workshop in Pittsburgh asks how robots can combine reusable skills and sensing modules as models grow larger. Topics include learning from scarce touch data, transferring skills between robot bodies and assembling longer tasks.
The public page links accepted papers and lists a panel on which parts of robot systems should remain modular. It offers a focused reading list for understanding how these design choices affect learning and reuse.
The workshop runs from 8:30 a.m. to 12:30 p.m. in Pittsburgh. Its papers are workshop contributions, and the event has not yet taken place; the schedule is not evidence that the proposed methods work.
Read this first
Watch RAPID’s project walkthrough to see how one demonstration becomes a sequence of contact-based moves. Keep the playback label in view: the featured physical example runs at ten times its recorded speed.
Read the paper