The Pulse
RACE Cuts Robot Idle Time With Longer Action Chunks
A research team introduced RACE, a method that predicts transitions between robot subskills to help vision-language-action models execute longer action chunks. In a pick-and-place test, RACE achieved 66% success with 20-action chunks, compa
AI.info Team · Written with AI ·
On October 5, researchers submitted a paper describing RACE, a method for letting vision-language-action models execute longer stretches of robot motion without pausing for as many policy updates. In a real-robot pick-and-place test, RACE used 20-action chunks—four times the five-action baseline—and recorded 66% success across 50 trials. The five-action baseline succeeded in 36% of trials.
The work targets a practical snag in robot control: large policies can take long enough to generate their next commands that a robot finishes its current sequence and waits. Longer action chunks can reduce those pauses, but the robot also has to act farther into the future without a fresh observation. The authors’ central finding is that errors gather around transitions between manipulation subskills, such as moving from approaching an object to grasping it.
RACE predicts when a robot should switch tasks
RACE stands for Reliable Action-Chunk Extension. The method first makes an auxiliary one-step denoising pass through the action model, then uses the resulting features to estimate where subskill transitions fall within the upcoming chunk. It conditions the model’s full action-generation process on those predicted transition timings.
The team builds RACE on the flow-matching vision-language-action model π0.5. The method does not depend on people manually labeling every transition: it derives training targets from action demonstrations using change-point detection. The authors describe the targets as pseudo-labels, rather than human-annotated ground truth.
“To understand where this unreliability arises, we analyze action errors within long chunks and find that they concentrate around transitions between manipulation subskills, growing sharply with chunk length.”
Seonghoon Yu, Dongwon Kim, HyungRok Jung, Yoonjae Baek, Byung-kwan Lee, Suha Kwak and Jeany Son, authors of the paper, write.
The real-robot gains come with an important comparison
The robot experiment used 30 demonstrations of a pick-and-place task and tested each policy over 50 trials. Researchers report results averaged over successful trials: the five-action fine-tuned policy took 12.6 seconds and spent 2.11 seconds idle. At 20 actions per executed chunk, ordinary fine-tuning reached 48% success with 0.40 seconds of idle time; RACE reached 66% with 0.41 seconds idle.
That distinction matters. The roughly fivefold idle-time reduction compares the 20-action RACE run with the five-action baseline, not with fine-tuning at the same 20-action chunk length. At 20 actions, RACE’s idle time is essentially the same as the fine-tuned model’s, while its success rate is higher. Completion time also changes little: the paper reports 11.8 seconds for RACE versus 11.6 seconds for 20-action fine-tuning, because asynchronous execution already hides much of the policy’s inference delay.
Simulation results expose the limit of longer chunks
The paper also evaluates RACE on VLABench, RoboCasa-H50 and LIBERO, three simulated manipulation benchmarks. On VLABench, RACE’s average success rate at 20 executed actions per policy call is 33.0%, compared with 28.1% for fine-tuning at the same chunk length. The authors report that RACE remains competitive against other vision-language-action approaches at four times the baseline chunk length, while longer chunks reduce the number of policy calls.
Longer is not always better. The authors say RACE’s transition predictor misses some transitions, particularly in held-out demonstrations and longer chunks, and success falls below the baseline at 40 executed actions. Transition-aware predictions address errors around subskill switches; they do not eliminate the general risk of acting open-loop for longer without new observations.
Fewer pauses mean fewer chances to react
The paper’s test covers a controlled laboratory task, not a robot operating around people or unpredictable objects. Its authors warn that longer chunks give a robot fewer opportunities to respond to unexpected changes, including someone entering its workspace, and say deployment near people requires safeguards such as safety monitoring, collision checking and human supervision.
RACE’s result is therefore specific: on the tested pick-and-place task, it raises success over ordinary fine-tuning at the same 20-action chunk length, while the long-chunk setup sharply reduces idle time against the five-action baseline. The method’s performance still declines when chunks grow too long, and its transition labels remain automatically inferred rather than hand-verified.