Visuomotor policies can handle many robot manipulation tasks, but understanding their failures remains difficult. Unfamiliar conditions and small action errors can lead to failure, yet watching the robot alone tells us little about what went wrong inside the policy. We introduce MIND (Mechanistic Interpretability for Neural Diagnosis), a method that looks inside a trained policy to detect failures during execution and identify the internal concepts involved. MIND uses a sparse autoencoder to break down the policy's internal activations into features that can be studied individually. We give meaning to these features by linking their activity to the robot's position, movement, and surroundings, and by steering their activations to examine how they affect the robot's actions. From successful and failed executions, we learn how the features behave on the way to a failure. During execution, MIND flags deviations from these patterns and uses the meaningful features contributing most to the change to describe the problem. Experiments in simulation and on a physical robot show that changes in these internal features help detect and explain failures. By using the same features for detection and diagnosis, MIND connects failure signals to understandable changes inside the policy.
Figures & tables
Figure 1: MIND end to end. The policy runs untouched (1–2), and a sparse autoencoder turns one of its activations into a sparse code (3–5). Top: a basis feature is named only if a recorded fact holds almost every time it fires (6), and a clamp tests whether the policy acts through it (7). Bottom: the basis features form a record at every step (8), a linear alarm scores it (9), the score splits over the concepts that moved it (10), and a language model puts them into words (11). Frames are recorded rollouts; traces follow the toy example of this section.
Figure 2: Two recorded attempts at one task: frames at the marked steps and, beneath, the six concepts of the toy vector, shaded by how strongly each basis feature is on (timing recorded, shading illustrative). Top: success. Bottom: failure, where the policy represents the gripper closing early, never represents the gripper near the bowl or the bowl grasped, and switches the downward decision on and off. The alarm is the first step at which the score stays above the threshold for three frames.
Figure 3: Names that stay precise on held-out episodes, for decision (dark) and state (light) concepts. (a, b) OpenVLA at unseen initial states and on the unseen task. (c) LingBot-VLA, action expert, layer 22 . (d) π0.5 in LIBERO, both streams, layers 7 , 11 and 15 (10 is LIBERO-10). (e) π0.5 on the robot, backbone, layer 11 . No baseline names any concept in (a), (b) or (e); Tables 4 – 6 give every baseline.
Figure 4: (a) Frames at which named basis features fire hardest on held-out episodes, with the held-out precision and recall of each name (blue: decision; green: state). (b) An OpenVLA failure: the frame at the alarm, the failure score and the note written from the concept record alone.
OpenVLA
LingBot-VLA
π0.5
π0.5 , real robot
Monitor
Spatial
Object
Goal
lift_pot
click_bell
4 tasks
stove
cups
Dense
1.00 (2%)
0.91 (7%)
0.92 (7%)
0.60 (4%)
0.63 (3%)
0.988 (6%)
0.74 (13%)
0.61 (13%)
Ours
1.00 (3%)
0.89 (5%)
0.92 (5%)
0.58 (4%)
0.72 (2%)
0.991 (3%)
0.70 (13%)
0.60 (0%)
Ours †
1.00 (2%)
0.93 (9%)
0.97 (2%)
–
–
–
0.84 (13%)
0.70 (13%)
Table 1: Failure detection: AUROC over held-out episodes, with false alarms (share of successful episodes) in brackets. Dense: monitor on the full activation; Ours: on the named basis features; † best variant fixed before any fit. The robot is scored leaving one attempt out. Table 8 gives intervals and variants; best AUROC in bold.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
meaning
value
dwell window lengths k , in frames
{4,16,64}
switching-rate window, in frames
16
share of successful episodes that fixes the onset deadline
0.9
false-alarm fraction on successful fitting episodes that sets the threshold
0.05
consecutive frames above the threshold before an alarm
3
Appendix
Table 2: Constants of the record and the alarm.
symbol
meaning
value
T
firing levels tried, as fractions of αimax
{0,0.1,0.2,0.3}
pmin
precision floor
0.9
rmin
recall floor of a named basis feature
0.05
nmin
minimum number of firings
50
–
minimum number of episodes a holding name fires in
5
–
largest share of those firings from one episode
0.5
Appendix
Table 3: Constants of the naming rule.
Figure 5: How g3 of the toy vector is named by the rule of Section 2.3 . (a) Naming frames. Each bar is the activation α3 on one frame, the dashed line is the firing level τα3max , and the basis feature fires where a bar rises above it. The row beneath marks the frames on which the concept, the gripper closing next, is true. The red bar is a firing on a frame where the concept is false, and it lowers the precision P . The circled frame is one where the concept is true but the basis feature is off, and it lowers the recall R . (b) Candidates. Of the concepts g3 fires with, only the gripper closing next clears the precision floor pmin , and it becomes the name. (c) Held-out frames. Precision and recall are measured again with the name and its level fixed, and the name holds.
Figure 6: Eight concepts that a basis feature of the LIBERO-Spatial dictionary earns a name for, each beside the frame at which that basis feature fires hardest over the recorded held-out episodes. Precision and recall are the ones the naming rule of Section 2.3 measures on episodes that took no part in choosing the name.
Figure 7: Four basis features of the dictionary at layer twenty, each named for the object in an image patch. In each pair, left is the set of patches on which the basis feature fires on a held-out frame, and right is where the simulator puts that object.
lift_pot
click_bell
Units
Decision
State
Decision
State
Dictionary, start 1 (ours)
28
9
11
5
Dictionary, start 2 (ours)
29
10
19
4
Raw coordinates, positive
25
4
9
3
Raw coordinates, negative
18
4
11
3
Random directions
10
4
10
2
Appendix
Table 4: Concepts of LingBot-VLA named on held-out attempts, at layer 22 of its action expert. Each task has 74 concepts, taken for each arm. The largest count of each column is in bold.
Units
LIBERO-10
Goal
Object
Spatial
Dictionary (ours)
53
48
42
50
of which decisions
17
16
16
15
of which states
36
32
26
35
Raw coordinates
16
14
16
13
Random directions
11
8
8
8
Named by the dictionary alone
36
32
28
37
Appendix
Table 5: Concepts of π0.5 named on held-out layouts, over both streams and the layers 7 , 11 and 15 .
Figure 8: Eight concepts that a basis feature of π0.5 on the physical robot earns a name for, read at the last token of the instruction in its backbone at layer 11 , each beside the camera frame at which that basis feature fires hardest over the held-out attempts of its task. The four decisions about the next action are all that the dictionary names there; the four states are those named with the highest recall, one per concept. Precision and recall are measured on attempts that took no part in choosing the name.
concepts of one token
every patch or token
Task
Units
B, state
B, next action
AE, last
AE, first
B
AE, last
stove
Dictionary (ours)
24 ( 11 )
4 ( 3 )
5 ( 5 )
0
122 (11)
61 (22)
Raw coordinates
0
0
1 (1)
7 ( 4 )
2 (2)
942 (23)
Random directions
0
0
0
0
78 (4)
3,231 (21)
Shuffled labels
0
0
0
0
0
0
Dictionary vs coord.
14 vs 0
5 vs 1
0 vs 4
Appendix
Table 6: Units of π0.5 on the physical robot named on held-out attempts, with the number of concepts they cover in brackets, at layer 11 . B: the backbone at the last token of the instruction; AE: the action expert at its last and first denoising steps; coord.: raw coordinates. The last two columns count the concepts true at every patch (backbone) or every action token (action expert, last step).
Figure 9: Clamping one basis feature of OpenVLA in closed-loop rollouts on the unseen task, from six to eight held-out initial states per suite. Each box is one clamped concept. Its upper bars give the share of steps whose action lies on the concept’s side, with no clamp, with the concept’s basis feature clamped, and under the two controls. Its lower bar gives how far the clamped arm ends from the unclamped one along the concept’s axis after twenty steps, and the turn has no such distance. Table 7 lists every value.
Figure 10: Clamping one named basis feature of OpenVLA in closed loop, on the held-out task from its first initial state. Each row starts from the same frame; without the clamp the policy follows its own course, and with the clamp (blue frames) the arm moves the way the concept says. Shares of steps are averaged over six to eight initial states (Table 7 ).
null controls
directions
suite
concept
no clamp
one basis feature
other basis features
random
summed
fitted
arm moves
Spatial
move +x
0.25
1.00
0.25
0.25
1.00
1.00
14 cm
Spatial
move −x
0.08
0.99
0.03
0.07
0.20
0.97
21 cm
Spatial
move −y
0.17
0.94
0.14
0.16
–
1.00
5 cm
Spatial
move down
0.34
0.68
0.33
0.40
1.00
0.97
10 cm
Object
move +x
0.16
1.00
0.21
0.14
0.96
1.00
14 cm
Appendix
Table 7: The values drawn in Figure 9 . Closed-loop rollouts on the unseen task, from six to eight held-out initial states. Each number is the share of steps whose action lies on the named concept’s side. The summed direction adds the named basis features of the concept, and the fitted direction is fitted to the concept’s labels by regression; a dash marks a concept for which no summed direction was formed. The last column is how far the arm clamped on one basis feature ends from the unclamped one along the concept’s axis after twenty steps.
Figure 11: The sign inversion of π0.5 in LIBERO, at layer 11 of the action expert. (a) Every direction chosen by prediction for +x points against the gradient of the action, while basis features labelled −x below the precision floor point along it. (b) Clamping the basis feature named for +x lowers the share of steps along +x , and clamping one labelled −x raises it to 1.00 ; each random bar is the control of the clamp to its left.
Task
Monitor
AUROC ↑
False alarms ↓
OpenVLA
Spatial
Dense activation
1.00 [1.00, 1.00]
2%
Ours
1.00 [0.99, 1.00]
3%
Ours, fused
1.00 [0.99, 1.00]
2%
Object
Dense activation
0.91 [0.81, 0.98]
7%
Ours
0.89 [0.78, 0.96]
5%
Ours, fused
0.93 [0.87, 0.98]
9%
Appendix
Table 8: Failure detection in full, with the intervals behind Table 1 . AUROC is per episode, with its 95% interval; false alarms are the share of successful episodes with an alarm. Dense activation: the same monitor on the activation itself. Ours reads the named basis features; fused takes the largest of their z-scored linear, per-position and tree scores. LingBot-VLA is scored on its opening frames, π0.5 in LIBERO over both streams at layers 7 , 11 and 15 , and the robot at layer 11 of its backbone, leaving one attempt out ( 8 successes). † Best robot monitor fixed before any fit (Appendix E.4 ). Best AUROC per task in bold.
Figure 12: Four failures that MIND called, one to a row: the camera frame at the step the alarm fired, the failure score over the whole attempt, and the sentence the language model wrote from the concept record alone. The grey line marks the step at which the simulator says the attempt has gone wrong, and the dot marks the alarm on the score.
Ablation
Policy
Measure
MIND
Ablated
Untrained policy
OpenVLA
names that hold
15–31 / 13–28
0 / 0
π0.5 , robot
probe AUROC, action expert
0.997–0.999
0.51
Neurons for the dictionary
OpenVLA
names that hold
15–31 / 13–28
0 / 0
π0.5 , LIBERO
one-step decisions named
6
1
Record without memory
OpenVLA
AUROC, Object / Goal
0.89 / 0.92
0.81 / 0.79
Length rule for the monitor
OpenVLA, π0.5
delay after onset
10–28
49–158
Appendix
Table 9: Ablations. Each row replaces one part of MIND and keeps the rest unchanged. Names are those that hold at unseen initial states, as decision / state basis features; delays are the median steps from the failure onset to the alarm.
Figure 13: The ablations of Table 9 . (a) Names that hold at unseen initial states, decision and state together. (b) Median probe AUROC over the stove concepts on the physical robot. (c) Detection AUROC with and without the dwells and the switching rate. (d) Median steps from the failure onset to the alarm, where lower is earlier.
Institute for AI Industry Research (AIR), Tsinghua University · University of Electronic Science and Technology of China · Work done during internship at AIR, Tsinghua University. +1