Robot learning

1.What does behavior cloning reduce robot policy learning to?

2.A kitchen robot learns to grab a mug. One operator always approaches from the left of the handle and another from the right. A policy trained with mean squared error drives straight into the handle. What is the most likely cause?

3.A lab pipetting arm has a behavior cloning policy trained on 50 demos. Training loss keeps falling, while loss on held-out demos rises after epoch 30. How should you pick the checkpoint to deploy?

4.Which assumption of plain behavior cloning breaks once the learned policy starts acting on its own?

5.A towel-folding robot sometimes freezes for several seconds mid-task. The demonstrations include long pauses where operators stopped to think. What is the best fix?

6.What is the main benefit of conditioning a behavior cloning policy on a language instruction or goal image?

7.A hospital delivery robot learns to press an elevator call button from a single current camera frame. The button's light is hard to see, and the robot keeps pressing it again and again. What is the most likely cause?

8.Why do small errors in a behavior cloning policy tend to compound over a long task?

9.What does DAgger-style training (dataset aggregation) add over plain behavior cloning?

10.An orchard drone flies between tree rows with a behavior cloning policy. Every demo was flown in the exact center of the row. After about 10 seconds the drone drifts toward the trees. What data fix helps most?

11.People find it hard to label the right action while watching a robot they aren't controlling. What practical variant of DAgger avoids this?

12.A lawn mower robot logs every operator takeover. An engineer adds the 3 seconds before each takeover to the imitation dataset as normal training data. What is the problem?

13.During teleoperation, the team adds small random perturbations to the robot's motion while the operator keeps controlling it. Why?

14.A behavior cloning policy runs at 10 Hz on a 20-second shelf restocking task. Assume each step has an independent 1 percent chance of an unrecoverable error. About what success rate should you expect?

15.An arm sorting library returns improves through weekly operator takeovers. Which metric best tracks progress between full evaluations?

16.Two operators collect demos. One is fast and smooth, the other slow with many retries. The policy trained on both behaves inconsistently. What do you do first?

17.Why should demonstrations be recorded through the same sensors and action interface the policy will use?

18.A pharmacy shelf-picking arm is taught by an operator standing next to the shelf and watching it directly. The policy sees only a wrist camera and fails on items the wrist camera can't see at the start. What is the core problem?

19.A team needs 400 demonstrations of a dishwasher loading task. Each demo takes 45 seconds and each reset takes 30 seconds. How much operator time is that, ignoring breaks?

20.Which teleoperation problem most directly hurts demo quality for a fine insertion task?

21.A sorting arm at a parcel hub logs camera frames on one computer and actions on another. Their clocks are not synchronized and drift apart by up to half a second per session. Policies trained on this data grasp late. What is the core problem?

22.A team has 500 demos, all on the same table with the same three mugs. What adds the most value for generalization next?

23.Why do many teams prefer a leader arm, a matching arm the operator moves by hand, over a gamepad for collecting manipulation demos?

24.A dataset of 2,000 teleoperated episodes for a pallet-wrapping robot has no success labels, and some episodes failed. What should you do before training a behavior cloning policy?

25.What is the main trade-off of a sparse reward, such as 1 only when a peg is fully inserted?

26.In simulation, a bin-picking arm earns reward each step for its gripper being close to the target item, plus a large bonus for lifting it. After training it hovers next to the item and never grasps. What is happening?

27.Why is reward shaping of the form gamma times phi(next state) minus phi(state) considered safe?

28.A simulated quadruped trained to walk fast gets high reward by vibrating its feet so it slides along the ground. On the real robot it barely moves. What is the best first diagnosis?

29.A reinforcement learning reward curve for a robot task is rising steadily. What should you do before trusting the trained policy?

30.A drawer-opening task runs at 20 Hz, takes about 30 seconds, and gives reward only at the end. With a discount factor of 0.99, about how much is that final reward worth from the first step?

31.A team uses a camera-based success classifier as the reward for learning to close a cabinet door. The policy learns to push the door nearly shut and hold the arm in front of the gap. What happened?

32.Why is sample efficiency a bigger concern for reinforcement learning on real robots than in simulation?

33.At a concept level, what makes an off-policy reinforcement learning method better suited to learning on a real robot than an on-policy one?

34.A team trains a legged locomotion policy with an on-policy method, using thousands of parallel simulated robots on one GPU. Why does on-policy training work well here despite being sample hungry?

35.A GPU simulator runs 4,096 robots in parallel, and each advances 50 policy steps per second of wall time. Roughly how long does it take to collect 1 billion steps?

36.A real arm learns a pushing task by reinforcement learning. Each episode lasts 20 seconds, and an automatic reset takes 25 seconds. How many episodes fit in an 8-hour shift with no downtime?

37.Reinforcement learning from scratch on a real cable-routing task makes almost no progress after a full day. You have 40 good demonstrations. What is a sensible next step?

38.What is the main extra risk when learning a robot policy by reinforcement learning purely from a fixed dataset, with no new interaction?

39.What does domain randomization aim to achieve in robot learning?

40.A team randomizes object mass from 0.01 kg to 20 kg for a policy that will only lift boxes of 0.2 to 1 kg. The policy turns slow and timid on every box. What is the likely cause?

41.A fruit-picking policy for a packing shed conveyor works in simulation but fails under the shed's real lights. Which randomization is most likely to help?

42.Why do teams often randomize action delay in simulation, for example between 0 and 40 milliseconds?

43.You can measure your real arm's joint friction and link masses in a day of tests. How should that change your domain randomization?

44.Which part of a manipulation task is usually least accurate in a rigid-body simulator?

45.A team must choose between a fast simulator with simple rendering and a slow photorealistic one to train a camera-based policy. What is a common, sound approach?

46.In simulation, a teacher policy sees exact terrain heights and friction, and a student policy learns to imitate it from onboard sensors only. Why use this setup?

47.A quadruped inspecting a pipeline walks smoothly in simulation, but on hardware its legs jitter and the motors run hot. What is the most likely gap?

48.A pick-and-place policy trained in simulation fails on the real robot. What is the most useful first step?

49.A policy trained in simulation places items about 2 cm to the left on the real robot, every time. The motion otherwise looks right. What is the most likely cause?

50.After transfer from simulation, a policy reaches 60 percent success on hardware. Which approach directly uses real experience to close the rest of the gap?

51.A policy trained on clean simulated depth images fails on a real depth camera that returns holes on shiny and dark surfaces. What helps most?

52.A policy trained in simulation at 50 Hz can only run at 30 Hz on the robot because image processing is slow. What problem does this create?

53.Besides physics, which difference is a common part of the sim-to-real gap?

54.What is a main advantage of end-effector (Cartesian) actions over joint position actions for a manipulation policy?

55.A policy controls a 7-joint arm with a parallel gripper using end-effector delta actions: a 3D position change, a 3D rotation as axis-angle, and one gripper command. How many numbers is each action?

56.What does action chunking mean in a learned manipulation policy?

57.A policy runs inference at 5 Hz and the robot's controller takes commands at 50 Hz. At minimum, how many actions must each chunk contain so the robot never runs out of commands?

58.A recycling sorting arm executes 2-second action chunks fully before looking again. When a worker bumps the bin, the arm keeps moving to the old position. What change helps most?

59.Why do learned manipulation policies often output targets at 10 to 50 Hz while a low-level controller runs at 500 Hz or more?

60.A joint controller runs at 1 kHz and the learned policy sends a new target at 20 Hz. How many controller cycles run for each policy target?

61.A drawer-opening policy uses only a fixed overhead camera and often misses the moment the gripper reaches the handle. Which added input most directly helps?

62.A policy outputs small position changes (deltas) relative to the current gripper pose. In a chunk of 20 steps executed without new observations, what risk grows along the chunk?

63.Why do many robot policies avoid predicting rotations as Euler angles?

64.A policy predicts gripper commands between 0 and 1 and joint deltas of about 0.001 radians. Training mostly ignores the joint deltas. What is the standard fix?

65.Why are diffusion-based action heads popular for imitation learning on manipulation tasks?

66.At a concept level, what is a vision-language-action model?

67.Which claim about vision-language-action models is most accurate?

68.A transformer policy takes the last 10 camera frames and joint states as input instead of only the current ones. What new failure mode should you watch for?

69.Some policies split each action dimension into 256 bins and predict them as tokens. What is the main trade-off?

70.A team has only 100 demonstrations for a new manipulation task. Which choice for the image encoder usually helps most?

71.Which image augmentation is most often helpful for visual behavior cloning with a fixed camera?

72.A diffusion policy uses 100 denoising steps, each taking 4 ms on the robot's GPU. How long does inference take per action chunk?

73.A tabletop policy trained with red, blue, and green blocks fails on a yellow block of the same shape. What does this suggest?

74.You want to claim a grasping policy generalizes to new objects. How should you design the test?

75.A policy trained on clean tables fails when a coffee mug sits near the target object. What data change targets this?

76.A sidewalk delivery robot's learned obstacle policy was trained on summer data. In autumn, leaves on the path cause frequent stops. What is the likely issue?

77.All demos for a peg task had the peg holder inside one 10 cm square. In a test, the holder sits 30 cm away, and the arm reaches toward the old area. Why?

78.A language-conditioned policy handles "pick up the apple" but fails on "grab the apple." What is the most likely cause?

79.A policy succeeds in 45 of 50 real trials. Using the normal approximation, what is the rough 95 percent confidence interval on its success rate?

80.Policy A succeeds in 14 of 20 real trials and policy B in 17 of 20. What should you conclude?

81.A safety check passes in 60 of 60 trials with zero failures. Roughly which failure rates can you rule out at 95 percent confidence?

82.Your success-rate interval from 25 real trials is too wide. About how many trials do you need to cut its width in half?

83.Why should real-robot evaluations follow a written reset protocol, with object positions drawn from a fixed list?

84.You compare two policies on one robot over a day. Policy A runs all morning and policy B all afternoon. What is the flaw?

85.A 5-step assembly task succeeds fully in 10 percent of trials. Why also report how many steps each trial completed?

86.Why do careful teams run real-robot evaluations blind, so the operator doesn't know which policy is running?

87.A team ranks 20 checkpoints in simulation and tests only the top 3 on the robot. When is this a reasonable shortcut?

88.A home mobile manipulator is evaluated with 200 trials, all in one test kitchen. What is the main limit of the result?

89.Before evaluating a shoe-pairing robot, why write down exactly what counts as success?

90.Before letting a learning policy explore on a real arm, which safeguard matters most?

91.A team wraps its learned policy in a safety layer that clips each command to a safe set before execution. What is a known downside?

92.A policy learning to insert connectors on an electronics line keeps jamming and spiking contact forces. Which low-level change most reduces risk to the parts?

93.A hospital delivery robot will learn from staff corrections in real hallways. What should be in place before the learned policy drives?

94.Why is large random exploration noise, common in simulated reinforcement learning, a concern on hardware?

95.During on-robot training, the policy process freezes and the controller keeps applying the last command. What safeguard prevents a crash?

96.A team pools data from five different arm types to train one policy. What is a key challenge?

97.A multi-task dataset has 20,000 picking episodes and 300 drawer-opening episodes. The pooled policy fails at drawers. What is a reasonable first fix?

98.A team has 50,000 robot episodes gathered from many labs. Which curation step most often improves imitation results?

99.Two labs contribute data for the same arm model. One records rotation actions in the gripper frame and the other in the base frame. What happens if you merge without converting?

100.A robot records three 640 by 480 RGB cameras at 30 frames per second, uncompressed, with 1 byte per channel. About how much data is that per second?

101.When scaling up robot data collection, which choice usually improves generalization more per hour of effort?

102.A team splits its robot data into training and validation sets by individual frame at random. Validation loss looks excellent, but real success is poor. What is wrong with the split?

103.A conveyor moves parts at 0.4 m/s. From image capture to robot action takes 250 ms. If the policy ignores this delay, about how far off will its grasp be?

104.A team wants to run a large policy on a cloud GPU for a warehouse mobile robot with spotty wifi. What is the main risk?

105.Inference takes 150 ms and the robot executes action chunks. Why might a team start the next inference before the current chunk ends?

106.A policy must run at 10 Hz. Image transfer takes 15 ms, preprocessing takes 10 ms, and sending the command takes 5 ms. If these run one after another with inference, what is the most time inference can take?

107.An airport baggage robot tests two policies trained on the same data. The large one wins on slow tabletop tasks but loses at catching tossed bags. The large one runs at 3 Hz and the small one at 30 Hz. What is the most likely reason?

108.Which change usually cuts onboard inference latency of a large policy with the least retraining?