When a robot practices without human resets, every attempt still costs wall-clock time, actuator wear, and sometimes a person's attention, yet the schedulers that decide what to practice next treat attempts as free. We present an open, low-cost testbed for measuring these costs: a self-built SO-101 arm runs two demonstration-trained ACT policies that reset each other, with a vision-based success detector and a ledger that records the time, joint travel, and human interventions of every attempt. We formulate cost-aware practice scheduling, in which task selection is rewarded by measured competence gain net of embodied cost, and specify in advance the comparison that will test it. Its first two conditions, run on hardware for 51 minutes of robot time, show that a fixed retry limit did not reduce the share of execution time lost to failures (36%, against 27% for a feasibility rule; not significant) and saved no robot time. The logs suggest why: failures showed no sign of clustering, so a retry right after a failure was as likely to succeed as any other attempt. The ledger also shows that most of the retry rule's extra human time came from cubes knocked out of reach rather than from the rule itself, and a cost-aware rule replayed on the logs recovers the sensible choice without a hand-set limit. We also report lessons from building the system, including a coverage effect in which concentrating demonstrations in a smaller workspace raised success there from 8/20 to 10/10.
Figures & tables
Figure 2: A robot we printed and assembled ourselves, doing both tasks. (a) Every structural part of the SO-101, printed in-house before assembly. (b) Place : the arm starts at its home pose, grasps the cube in the taped square, and releases it over the tray. (c) Retrieve : the same arm reaches into the tray, grasps the cube, and sets it back in the square, which is exactly where place begins.
Figure 3: The practice loop, drawn with real frames from one unattended cycle. (1) The scene camera sees the cube, and the detector marks it as on the table. (2) The scheduler turns that into the next attempt: place from the table, retrieve from the tray, a call for help if the cube is out of reach. (3) The chosen ACT policy drives the arm. (4) The detector scores the result; here the cube has settled in the tray. (5) The ledger writes one row for the attempt; the values shown are a real row from our logs. (6) The arm returns to the pose every demonstration started from, and the loop repeats. The dashed arrow is what this paper adds: today’s scheduler records the cost but never reads it, and the proposed scheduler decides with it.
Condition
Action set
Reward
Role
A. Feasibility
attempt only
none
the pilot baseline
B. Fixed retry limit
attempt, request help
none
requests help after N consecutive failures, a fixed reset budget in the spirit of Leave No Trace
C. Cost-blind bandit
attempt, reposition, request help
success, λ=0
isolates the effect of a learned schedule
D. Cost-aware bandit
attempt, reposition, request help
success −λ⊤c
the proposed method
Table 1: Four schedulers on the same arm, from no choice to a priced choice. A and B are the rule-based baselines reported in Section 6.4 ; C adds a learned schedule without prices; D is the proposed method.
Metric
A. Feasibility
B. Retry limit ( N=2 )
Sessions, robot minutes
3, 30.9
2, 20.1
Valid attempts
47
32
Success rate (95% CI)
78.7% (65.1–88.0)
65.6% (48.3–79.6)
Place, retrieve successes
17/26, 20/21
11/21, 10/11
Execution time on failed attempts ∗
26.6%
35.7%
Successes per robot minute
1.20
1.05
Table 2: Conditions A and B, pooled over sessions. Confidence intervals are Wilson score intervals over attempts; they treat attempts as independent and are therefore optimistic. ∗ Execution time excludes homing and detection, which Figure 4 counts separately.
Figure 4: A quarter of the robot’s time goes to neither success nor failure. (a) How every minute of robot time was spent under each rule. Blue is useful work, orange is failed attempts, and gray is homing and waiting for the detector, an overhead paid on every attempt whichever rule chooses it. (b) What one success cost in each of the ledger’s three currencies. Robot time here includes homing, so it is higher than the execution time in Table 2 . Most of rule B’s human time came from cubes knocked out of reach, including a single 130 s repair, not from the retry rule itself ( Section 6.4 ).
Figure 5: Every attempt in all five sessions, in the order it happened. Each block is one attempt and its width is how long it took; blue succeeded and hatched orange failed. Black triangles mark the retry limit stopping to ask for help; gray triangles mark a person stepping in, either to set up the cube at the start of a session or to move one that was out of reach. The number at the right of each row is that session’s successes. Long runs of orange are rare, which is why a rule that counts failures had little to act on.
Figure 6: A failure does not predict another, and most failures are cheap. (a) A place attempt straight after a failed place succeeded 62% of the time, the same as place attempts overall; bars are 95% Wilson intervals. (b) The duration of every valid attempt, with black lines at the medians and the 800-step cap dashed. Most failed place attempts end early, because the gripper releases over the tray and the cube lands beside it; the expensive failures are the few that run to the cap.
Robot policies inevitably encounter failures when deployed in real environments. Naive retries often repeat the same mistakes, while many existing recovery methods rely on human intervention. In this paper, we propose Failure-Aware Retry (FAR), a framework that enables robots to learn from previous failures at test time, adapt their behavior accordingly, and eventually complete the task autonomously. FAR combines Failure-Contrastive Preference Adaptation, which constructs preference learning data from failures to steer the policy away from previously unsuccessful behaviors, with lightweight action perturbations during retries to encourage local exploration. We further incorporate successful recovery trajectories into a training loop for continual policy improvement. Experiments in both simulation and real-world manipulation tasks show that FAR substantially improves success rates and robustness, with average gains of 17.6% over the standard diffusion policy in simulation and 11.7% in the real world. In addition, FAR improves data efficiency under both reset and timestep budgets during continual policy improvement by exploiting informative failure cases. Videos and code are available at https://hoar012.github.io/FAR-Project.
We consider the problem of autonomously learning robot skills under a limited practice budget for sequential tasks. We propose an active skill learning algorithm, Deliberate Practice (DP), that computes a provably budget-optimal allocation---practicing skills that maximize expected cumulative reward while being learnable within the budget. DP estimates both the time needed to master skills and the cumulative reward of the task plans that the skills unlock. Computing a budget-optimal allocation is challenging as it requires reasoning about combinatorially many skill plans over a large practice budget. Our key contribution is a bilinear program that can compute this exactly using off-the-shelf solvers. Through simulated and real-world experiments on long-horizon manipulation tasks, we show that our approach allows robots to optimally use limited practice time to acquire useful policies and improve long-horizon planning.
Shivam Vats, Sudarshan Harithas, Mete Tuluhan Akbulut +2
1Brown University · 2Mitsubishi Electric Research Laboratories
The pace of a robot demonstration is often determined by the teleoperator rather than by the task itself. As a result, the same motion can appear at widely different speeds across demonstrations, and policies trained on such data may inherit operator-specific timing along with task behavior. We propose to learn a speed prior directly from large-scale demonstration data. Given a task instruction, a motion path and the images along it, it predicts a distribution over execution speed along the path, from which a chosen quantile defines a consistent clock for retiming each demonstration. Existing methods derive the acceleration signal from the very data they accelerate, so each new task costs another fit, annotation, or pair of predefined rates. A single trained speed prior instead covers every task in a corpus. We fit a speed prior on both robot teleoperation data and handheld demonstrations in which a human arm, rather than a robot arm, moves the gripper, and use it to retime robot demonstrations to the handheld pace before fine-tuning a vision-language-action policy. For execution, we pair the policy with a real-time optimal-control layer that tracks predicted action chunks under robot constraints using the predicted pace. On 120 RoboLab tasks, policies fine-tuned on our learned clocks succeed more often and move with up to half the jerk compared with policies fine-tuned on the original demonstration timing and uniformly accelerated to the same speed. On a real robot, our policy stacks cups in 3.50s, 1.72× faster than the default policy's 6.02s while succeeding as often, in 29 of 30 trials. Videos and further material are available at https://mllab.snu.ac.kr/projects/learned_speed_prior.