acceptodds
Under review as a conference paper at ICLR 2027

SURE: Support-Conditioned Unreliability-Aware Residual Exteroception for Graceful Degradation in Humanoid Locomotion

Abstract

Visual terrain perception helps humanoids traverse complex terrain, but erroneous or stale elevation maps can misdirect foot placement and whole-body motion. Controllers therefore need to determine when to stop relying on map guidance while maintaining basic locomotion after a mismatch. We propose SURE (Support-Conditioned Unreliability-Aware Residual Exteroception), a visual-residual control method based on physical consistency between vision and proprioception. SURE adds a visual residual to a blind base policy driven by proprioception, providing terrain-dependent action corrections. To decide when to use these corrections, it infers foot support and collision states from proprioception and checks them against local map geometry. Vertical checks compare supporting-foot height with mapped terrain height, while horizontal checks compare foot-front obstruction with mapped obstacle geometry. Physical contradictions prompt residual withdrawal, leaving the blind base policy in control. Subsequent foot placements must pass support verification before visual corrections resume. Simulations show that SURE preserves complex-terrain locomotion under nominal vision. Under local geometric mismatch and map delay, SURE improves walking success relative to the same policy with the visual residual always enabled. Restoration ablations show that map-based support verification improves completion beyond proprioceptive recontact under delay and height drift. Real-robot experiments further demonstrate terrain traversal with omitted obstacles or delayed maps. These results show that physically grounded residual control reconciles terrain adaptation under nominal vision with basic locomotion after perceptual mismatch, providing interpretable graceful degradation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
humanoid locomotion, perceptive locomotion, residual control, cross-modal physical consistency, graceful degradation, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8LcPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces SURE, a humanoid locomotion controller for unreliable terrain maps. SURE keeps a blind policy as the base and adds a visual residual when vision is used. The controller compares estimated foot support and collision events with the map to decide when to withdraw the visual residual and when to use it again. The experiments in paper reports improved robustness under map faults in simulation environments, it also shows deployment experiments on real robot.

Critical strengths

The described problem is practical, as the robot cannot always trust its terrain map. It also makes lot sense to treat withdrawing vision and restoring vision as two separate decisions. Getting foot contact again doesn't mean the map is correct. The blind base stays active, so withdrawing vision doesn't require switching to a separately trained policy. These points claimed are reasonable

Critical weaknesses
  1. I see this as a useful but fairly incremental improvement in vision-blind fallback. The physical checks provide a concrete way to withdraw and restore visual corrections. My concern is whether this design brings enough additional value over existing fallback methods, beyond using a different rule to decide when vision should be trusted.

  2. Vision-blind fallback is already explored in previous works, for example: VB-Com, which the paper already cited, in VB-Com, the decision uses predicted returns. SURE instead uses physical consistency. The question is what additional practical value does this physical-consistency criterion bring over return-based switching?

  3. I am not sure how much of the gain over HS comes from the new design, since the training is also different. SURE trains the blind and visual parts together, while HS trains them separately.

  4. Hardware side, the tests only show the proposed method. What happens if the robot keeps using vision in these same situations? Successful runs show that SURE can work, but don't really tell us how much the proposed design improves reliability on real robots

Questions for authors

Please see weakness, questions are mentioned in weakness above

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8LcP (1/4)

Thank you for asking what SURE adds beyond existing fallback methods and how the selection criterion and action structure each affect performance.

W1. Contribution beyond existing fallback

Answer to W1: We would like to clarify that SURE's main contribution is to use cross-modal verification to decide whether to apply a map-driven correction to the blind action. It compares the geometry given by the elevation map with proprioceptive estimates of support and collision at the same foot locations. The event estimator does not read the map, so it provides a separate source of evidence.

VB-Com selects between complete blind and visual actions using predicted returns. A low blind return can reflect a difficult step even when the map is correct. Miki et al. combine proprioception and exteroception in a learned belief. SURE instead keeps the two sources separate so that it can check the mapped geometry against estimated foot interactions.

Both SURE and VB-Com can revert to blind control. In SURE, even with vision enabled, the visual branch takes the blind action as input and learns a correction to it. At a given observation, withdrawing vision removes that correction without changing the underlying blind action.

The design therefore changes both the evidence used to withdraw visual guidance and how vision contributes to the action while it is enabled.

AuthorsOpenReview ↗

Response to Reviewer 8LcP (2/4)

W2. Practical value over return-based switching

Answer to W2: The practical gain is retaining visual assistance on difficult terrain when the map is correct. On the same frozen controller, SURE achieved 100% on high-step ascent (T2-up) and 99.40% on repeated gutters (T5). VB-Com-style return selection achieved 0% on both.

To compare the criteria directly, we tested both on identical terrains and faults. Each selects between the same blind base (B) and visually assisted mode (B+R), where R is the visual residual.

Return selection predicts locomotion performance from proprioceptive history. SURE instead checks the received map against map-independent estimates of support and collision at the corresponding feet.

The return heads use five-frame proprioceptive histories and additional clean and fault-matched rollouts covering G1–G4. Independent critics with training-only privileged inputs provide advantage-plus-value (GAE+V) targets averaged over 50 steps. On independent validation, the B/B+R heads explain most of the target variation (explained variance 0.83/0.85). Both criteria use the direct stream-loss alarm for G5.

Success rate (%). Parentheses give evaluation-seed SD in units of the mean's last digit. G1–G5 denote unmapped protrusions, false support, map delay, height drift, and stream loss.

ConfigurationCleanG1G2G3G4G5
SURE99.93 (2)100.00 (0)100.00 (0)95.60 (56)100.00 (0)100.00 (0)
Return selection77.66 (3)100.00 (0)100.00 (0)88.80 (61)100.00 (0)100.00 (0)

Return selection handles several faults well. Its clean-terrain failures show that a criterion can reject misleading guidance yet also withdraw vision needed for traversal.

To test the rules without learned prediction error, we keep both rules and replace their scores with measured finite-horizon returns. Simulating both modes from cloned decision states with their map inputs is computationally expensive, so we use 10 episodes per scene. T2-up remains 0/10 and T5 1/10.

  • Relative-return comparison. In T5, some withdrawals occur even after the current visual return exceeds the blind return. Earlier low visual scores remain in the averaging window, because the rule compares a smoothed visual score with the current blind score. Those different time references can cause withdrawal on a correct map.
  • Blind-return decline. This rule requests blind control when the current blind score is at or below its recent mean minus a margin. In T2-up, the decline rule still requests blind control when the visual score exceeds the current blind score both before and after smoothing. A falling blind score on difficult terrain can indicate a greater need for vision, not a map error.

The recorded decisions explain why better prediction alone need not resolve these failures. SURE's consistency criterion withdraws the visual residual when the map contradicts physical evidence, not simply when locomotion becomes harder.

AuthorsOpenReview ↗

Response to Reviewer 8LcP (3/4)

W3. Separating action structure from training

Answer to W3: To separate action structure from training, we compared the residual design with jointly trained full-action heads. SURE retains its success-rate advantage on unmapped protrusions and map delay.

The comparison removes the separate-versus-joint training difference in the original hard-switching baseline. We matched the full-action heads to SURE in terrain and visual-availability curricula, model capacity, and consistency rules. The full-action heads share proprioceptive features and use an event head calibrated on their own clean rollouts.

The full-action heads replace the blind action with a full visual action (a = (1−β)a_B + βa_V). SURE instead adds a visual residual to the proprioceptive blind base (a = a_B + βa_R). Binary β enables or withdraws visual assistance.

Success rate in the joint-training structural comparison (%). Parentheses give evaluation-seed SD in units of the mean's last digit.

ConfigurationCleanG1G2G3G4G5
SURE99.92 (3)100.00 (0)100.00 (0)96.90 (40)100.00 (0)100.00 (0)
Full-action heads98.08 (33)0.00 (0)99.93 (12)79.23 (172)100.00 (0)99.90 (10)

G1 denotes unmapped protrusions, G2 false support, G3 map delay, G4 height drift, and G5 stream loss.

Both designs perform well on G2/G4/G5. SURE's success advantage is clearest on G1/G3 despite matching joint training.

We use observations collected during each method's own visually assisted rollouts. At each observation, we evaluate both modes and compare their targets for the 12 controlled leg joints. The mean RMS gap is 0.07114 versus 0.12107 rad (41.2% smaller for SURE). The mean within-episode 95th percentile is 0.12761 versus 0.20967 rad (39.1% smaller). This is a between-mode difference, not a temporal action difference. A smaller gap means the leg-joint targets change less if vision is withdrawn at the evaluated state.

Motion metrics. TT and WW are torso-tilt and torso-angular-velocity CVaR95; EvE_v is velocity-tracking error and DaD_a is time-normalized first-order action-difference RMS.

ConditionConfigurationTT (rad)EvE_v (m/s)DaD_a (s⁻¹)WW (rad/s)
CleanSURE0.073 (<1)0.088 (<1)14.947 (<1)1.766 (<1)
CleanFull-action heads0.096 (<1)0.092 (<1)16.858 (4)1.899 (1)
G1–G5SURE0.080 (<1)0.088 (<1)16.892 (3)2.133 (1)
G1–G5Full-action heads0.084 (<1)0.114 (<1)22.923 (20)2.329 (3)

Scene means include valid motion up to termination, including from failed episodes, and are equally weighted within each group. (<1) denotes a nonzero SD that rounds to zero.

In clean rollouts, SURE reduces first-order action-difference RMS by 11.3% and torso-tilt CVaR95 by 23.2%, with lower tracking error. Under faults, first-order action-difference RMS is 26.3% lower and tracking error falls from 0.114 to 0.088 m/s. The same-observation comparison shows why sharing a blind reference can reduce the action change at withdrawal. The rollout results show that the advantage extends to motion during traversal.

AuthorsOpenReview ↗

Response to Reviewer 8LcP (4/4)

W4. Always-visual control in the hardware stair task

Answer to W4: We reproduced the hardware stair task in an idealized, repeatable simulation to avoid further hardware damage. An earlier always-visual trial had resulted in an ankle-joint malfunction after a collision. With a delayed map, always-visual control rarely finished the route. The robot often retreated and advanced again, with frequent leg collisions.

The three configurations share one controller, with a persistent 500 ms map delay and a 0.5 m/s command. Success means reaching the endpoint within 20 s. Leg-collision rate is the share of episodes with leg–terrain collisions; those collisions do not end the episode.

MethodSuccess rate (%)Leg-collision rate (%)
SURE100.000 ± 0.0004.07
Always blind100.000 ± 0.0001.40
Always visual2.167 ± 0.55199.80

SURE and always-blind control completed every route, showing the blind base can traverse the stairs. Withdrawing misleading visual correction preserved route completion and reduced leg collisions relative to always-visual control.

Reviewer_S2akOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SURE addresses humanoid locomotion with unreliable elevation maps. It adds a visual residual to a blind locomotion policy and uses physical consistency checks to decide when to disable that residual. A proprioceptive event head estimates foot support and collisions, which are compared with the corresponding map geometry. Visual corrections resume only after subsequent foot placements provide evidence of support consistent with the map. The paper evaluates nominal locomotion, several map faults, withdrawal and restoration mechanisms, and deployment on a Unitree G1.

Critical strengths
  1. The fallback decision has a clear physical basis. Comparing map geometry with foot support and collision feedback makes it easy to understand why the controller withdraws visual corrections.

  2. The same-weight comparisons are useful controls. The always-blind and always-visual configurations help isolate the contribution of dynamic gating.

  3. The restoration ablation is convincing. Table 2 shows that checking agreement with the map improves completion beyond waiting for foot recontact alone. This supports using different criteria for withdrawal and restoration.

  4. The hardware experiments add practical evidence. The G1 trials demonstrate operation with obstacles omitted from the map and with delayed map updates. Although limited in scale, they provide a useful check that the approach works beyond simulation.

Critical weaknesses
  1. VB-Com is discussed as a closely related method but is not evaluated. Since hard switching uses the same family of consistency checks, it does not establish how the proposed criterion compares with return-based selection.

  2. The hard switching comparison mixes differences in policy composition and training. The shared-schedule experiment controls withdrawal timing, but the contribution of residual parameterization remains unclear.

  3. Both blind configurations achieve 100% success in every fault condition. The task-completion benefit of restoring vision after a fault therefore remains less thoroughly tested.

Questions for authors
  1. How does SURE compare with VB-Com under the same terrain and fault conditions? This would help clarify the advantages and tradeoffs of physical-consistency gating.

  2. Could a baseline with two jointly trained action heads, using the same curriculum and gating rule, help separate the benefit of residual parameterization from joint training? The shared-schedule result is useful, but it does not resolve this distinction.

  3. Could you evaluate a recoverable map fault followed by terrain that requires visual preview? Comparing SURE with permanent blind fallback on this sequence would directly test whether verified restoration recovers the capability needed to finish the task.

6 comments
AuthorsOpenReview ↗

Response to Reviewer S2ak (1/4)

Thank you for suggesting comparisons that distinguish the roles of the selection criterion, the residual action structure, and visual restoration.

W1. Missing return-based comparison

Answer to W1: We added VB-Com-style return selection on SURE's frozen controller. Unlike the original hard-switching baseline, which uses the same consistency checks, this compares different selection criteria with identical action policies (Q1).

W2. Separating action structure from training

Answer to W2: Jointly trained full-action heads remove the separate-versus-joint training difference that the shared switching schedule leaves unresolved. Q2 compares their action compatibility and motion quality with SURE's residual design.

W3. Testing the benefit of visual restoration

Answer to W3: We tested a recoverable map fault followed by terrain that requires visual preview. The comparison with permanent withdrawal uses the same trigger to test whether restoration helps the robot finish the route (Q3).

AuthorsOpenReview ↗

Response to Reviewer S2ak (2/4)

Q1. Return-based selection under the same terrain and faults

Answer to Q1: On the same frozen controller, SURE achieved 100% on correctly mapped high-step ascent (T2-up) and 99.40% on repeated gutters (T5). VB-Com-style return selection achieved 0% on both.

Following your suggestion, we evaluated both criteria on identical terrains and faults. Both select between the same blind base (B) and visually assisted mode (B+R), where R is the visual residual.

Return selection predicts locomotion performance from proprioceptive history. SURE instead checks the received map against map-independent estimates of support and collision at the corresponding feet.

The return heads use five-frame proprioceptive histories and additional clean and fault-matched rollouts covering G1–G4. Independent critics with training-only privileged inputs provide advantage-plus-value (GAE+V) targets averaged over 50 steps. On independent validation, the B/B+R heads explain most of the target variation (explained variance 0.83/0.85). Both criteria use the direct stream-loss alarm for G5.

Success rate (%). Parentheses give evaluation-seed SD in units of the mean's last digit. G1–G5 denote unmapped protrusions, false support, map delay, height drift, and stream loss.

ConfigurationCleanG1G2G3G4G5
SURE99.93 (2)100.00 (0)100.00 (0)95.60 (56)100.00 (0)100.00 (0)
Return selection77.66 (3)100.00 (0)100.00 (0)88.80 (61)100.00 (0)100.00 (0)

Return selection handles several faults well. Its clean-terrain failures show that a criterion can reject misleading guidance yet also withdraw vision needed for traversal.

To test the rules without learned prediction error, we keep both rules and replace their scores with measured finite-horizon returns. Simulating both modes from cloned decision states with their map inputs is computationally expensive, so we use 10 episodes per scene. T2-up remains 0/10 and T5 1/10.

  • Relative-return comparison. In T5, some withdrawals occur even after the current visual return exceeds the blind return. Earlier low visual scores remain in the averaging window, because the rule compares a smoothed visual score with the current blind score. Those different time references can cause withdrawal on a correct map.
  • Blind-return decline. This rule requests blind control when the current blind score is at or below its recent mean minus a margin. In T2-up, the decline rule still requests blind control when the visual score exceeds the current blind score both before and after smoothing. A falling blind score on difficult terrain can indicate a greater need for vision, not a map error.

The recorded decisions explain why better prediction alone need not resolve these failures. SURE's consistency criterion withdraws the visual residual when the map contradicts physical evidence, not simply when locomotion becomes harder.

AuthorsOpenReview ↗

Response to Reviewer S2ak (3/4)

Q2. Residual parameterization under joint training

Answer to Q2: To address the training difference you pointed out, we compared the residual design with jointly trained full-action heads. SURE retains its success-rate advantage on unmapped protrusions and map delay.

The comparison removes the separate-versus-joint training difference in the original hard-switching baseline. We matched the full-action heads to SURE in terrain and visual-availability curricula, model capacity, and consistency rules. The full-action heads share proprioceptive features and use an event head calibrated on their own clean rollouts.

The full-action heads replace the blind action with a full visual action (a = (1−β)a_B + βa_V). SURE instead adds a visual residual to the proprioceptive blind base (a = a_B + βa_R). Binary β enables or withdraws visual assistance.

Success rate in the joint-training structural comparison (%). Parentheses give evaluation-seed SD in units of the mean's last digit.

ConfigurationCleanG1G2G3G4G5
SURE99.92 (3)100.00 (0)100.00 (0)96.90 (40)100.00 (0)100.00 (0)
Full-action heads98.08 (33)0.00 (0)99.93 (12)79.23 (172)100.00 (0)99.90 (10)

G1: unmapped protrusions; G2: false support; G3: map delay; G4: height drift; G5: stream-loss alarm.

The full-action comparator performs well on clean terrain; the largest success gaps are on G1/G3.

We use observations collected during each method's own visually assisted rollouts. At each observation, we evaluate both modes and compare their targets for the 12 controlled leg joints. The mean RMS gap is 0.07114 versus 0.12107 rad (41.2% smaller for SURE). The mean within-episode 95th percentile is 0.12761 versus 0.20967 rad (39.1% smaller). This is a between-mode difference, not a temporal action difference. A smaller gap means the leg-joint targets change less if vision is withdrawn at the evaluated state.

Motion metrics. TT and WW are torso-tilt and torso-angular-velocity CVaR95; EvE_v is velocity-tracking error and DaD_a is time-normalized first-order action-difference RMS.

ConditionConfigurationTT (rad)EvE_v (m/s)DaD_a (s⁻¹)WW (rad/s)
CleanSURE0.073 (<1)0.088 (<1)14.947 (<1)1.766 (<1)
CleanFull-action heads0.096 (<1)0.092 (<1)16.858 (4)1.899 (1)
G1–G5SURE0.080 (<1)0.088 (<1)16.892 (3)2.133 (1)
G1–G5Full-action heads0.084 (<1)0.114 (<1)22.923 (20)2.329 (3)

Scene means include valid motion up to termination, including from failed episodes, and are equally weighted within each group. (<1) denotes a nonzero SD that rounds to zero.

In clean rollouts, SURE reduces first-order action-difference RMS by 11.3% and torso-tilt CVaR95 by 23.2%, with lower tracking error. Under faults, first-order action-difference RMS is 26.3% lower and tracking error falls from 0.114 to 0.088 m/s. The same-observation comparison shows why sharing a blind reference can reduce the action change at withdrawal. The rollout results show that the advantage extends to motion during traversal.

Reviewer_r2MfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SURE, adds a visual residual on top of a blind, proprioception-only locomotion policy. Instead of a naive residual setting, the authors propose a global binary switch param β that turns the residual on or off. A GRU event head estimates each foot’s support and collision state from proprioception. Two hand-designed checks compare these states with the elevation map: Vertical check and horizontal check. in simulation, SURE keeps nominal success (99.93%) and reaches 96–100% under five map faults. Ablations isolate the two checks and the restoration rule. A Unitree G1 completes 4 conditions × 5 trials.

Critical strengths

First, the paper proposes a simple, interpretable mechanism. Each withdrawal can be traced to a specific foot and type of contradiction.

Second, a nice comparison setting, β = 0 and β = 1 share the same policy weights, which isolates the effect of the gating. The shared-schedule experiment shows the advantage over hard switching is not only about when the switch happens.

Third, a useful finding on restoration. Withdrawal and restoration need different evidence. Re-contact alone confirms the foot is supported, not that the map is correct. Map-verified restoration raises G4 success from 92% to 100%, and the ablation shows this cleanly.

Critical weaknesses
  1. The proposed method's novelty is limited for the ICLR community. The learned parts are standard: a residual policy trained with PPO and a supervised contact/collision estimator. The main contribution fits better to a robotics conference, where there is an incremental learning-based method that solves a real robotics problem.

  2. The randomization is too little. The visual baseline is trained with only ±3 cm map noise, so its 0% success under faults is expected. Heavy-DR’s randomization does not match the tested fault types. In addition, at least the paper should compare against VB-Com, the closest prior method, which is cited but not compared.

  3. The choice of the binary, global, permanent fallback might be a limiting design. A single local inconsistency removes all visual guidance. Under persistent drift (G4), the residual stays off for most of the run (18% visual-on), so the robot effectively becomes blind. It is unclear why the method discards the map rather than re-registering or locally masking it using the contact evidence it already has. Or alternative, can this design choice offer such a capability?

Questions for authors

Question 1: How does SURE perform when map faults occur on terrain the blind policy cannot traverse, such as high steps or gutters?

Question 2: Can you compare against VB-Com and a Miki-style belief encoder, and against a visual policy trained with the tested fault types?

4 comments
AuthorsOpenReview ↗

Response to Reviewer r2Mf (1/4)

Thank you for the careful review and for asking us to clarify SURE's contribution beyond standard learning components and its rationale for global withdrawal.

W1. Contribution beyond standard learning components

Answer to W1: We would like to clarify that SURE's main contribution is to use cross-modal verification to decide whether to apply a map-driven correction to the blind action. It compares the geometry given by the elevation map with proprioceptive estimates of support and collision at the same foot locations. The event estimator does not read the map, so it provides a separate source of evidence.

VB-Com selects between complete blind and visual actions using predicted returns. A low blind return can reflect a difficult step even when the map is correct. Miki et al. combine proprioception and exteroception in a learned belief. SURE instead keeps the two sources separate so that it can check the mapped geometry against estimated foot interactions.

Both SURE and VB-Com can revert to blind control. In SURE, even with vision enabled, the visual branch takes the blind action as input and learns a correction to it. At a given observation, withdrawing vision removes that correction without changing the underlying blind action.

PPO and supervised learning train the policy and the event estimator, respectively. The methodological contribution is to connect this cross-modal check to a visual correction that can be withdrawn without replacing the blind action.

W2. Fault-matched randomization and baselines

Answer to W2: We added fault-category-matched domain randomization (Matched-DR Visual) beyond nominal ±3 cm noise. Training covers G1–G4: omitted protrusions, filled depressions, map delay, and height drift. It reaches 99.95% success on clean terrain. Q2 compares SURE with all three added baselines: Matched-DR Visual, VB-Com-style return selection, and Miki-style belief adaptation.

AuthorsOpenReview ↗

Response to Reviewer r2Mf (2/4)

W3. Why binary, global withdrawal instead of map repair?

Answer to W3: SURE uses binary, whole-residual withdrawal because a map mismatch does not tell us how to repair the map or which part of the correction to keep. Incorrect geometry can misguide foot clearance or placement. The residual coordinates whole-body actions from fused map and proprioceptive features and is not decomposed by map region. Setting β to zero removes the correction and leaves the blind base in control without discarding the map.

Why not repair or mask the map? Re-registration can address pose error but cannot reconstruct missed obstacles or pits. Local masking requires knowing the affected region and what should replace its input. Withdrawal removes the effect of erroneous geometry on actions without waiting for either; map repair can complement it.

Consistency-triggered withdrawal is reversible. The map remains available for checks. Visual corrections resume when new foot support agrees with the map and the restoration conditions are met. The low visual-on fraction under persistent drift reflects continued mismatch, not a permanent lock to blind control.

Q1. Faults on terrain beyond blind capability

Answer to Q1: If a map fault persists on terrain beyond blind capability, withdrawal stops erroneous geometry from guiding the action but cannot supply the missing traversal capability. Continuing to use the faulty map would not provide reliable visual preview either.

AuthorsOpenReview ↗

Response to Reviewer r2Mf (3/4)

Q2. Comparison with fault-matched baselines

Answer to Q2: SURE retains near-perfect clean success and matches or exceeds the three comparators on the tested faults.

Following your suggestion, we added three fault-matched comparisons. VB-Com-style return selection and Miki-style belief adaptation train their added modules on SURE's frozen controller. B is the proprioceptive blind base, and B+R adds the visual residual R. Matched-DR Visual is trained separately. All three comparators were trained with G1–G4 fault types.

Success rate (%). Parentheses give evaluation-seed SD in units of the mean's last digit. G1–G5 denote unmapped protrusions, false support, map delay, height drift, and stream loss.

ConfigurationCleanG1G2G3G4G5
SURE99.93 (2)100.00 (0)100.00 (0)95.60 (56)100.00 (0)100.00 (0)
Return selection77.66 (3)100.00 (0)100.00 (0)88.80 (61)100.00 (0)100.00 (0)
Belief adaptation99.98 (1)100.00 (0)0.00 (0)82.23 (67)100.00 (0)90.53 (12)
Matched-DR Visual99.95 (1)0.00 (0)100.00 (0)74.20 (62)100.00 (0)0.00 (0)

In G5, a stream-loss alarm triggers blind control for SURE and return selection. Belief adaptation supplies adapted features to B+R, while Matched-DR Visual has no blind mode. Both receive the fallback map without alarm-based switching. G5 therefore tests stream-loss handling, not consistency detection alone.

Q2(a). VB-Com-style return-based selection

Answer to Q2(a): SURE retains the visual assistance needed on difficult terrain when the map is correct. Return selection handles several faults well, but on the same frozen controller, SURE achieves 100% on high-step ascent (T2-up) and 99.40% on repeated gutters (T5), versus 0% on both for return selection.

The return heads use five-frame proprioceptive histories and additional clean and fault-matched rollouts. Independent critics with training-only privileged inputs provide advantage-plus-value (GAE+V) targets averaged over 50 steps. On independent validation, the B/B+R heads explain most of the target variation (explained variance 0.83/0.85).

To test the rules without learned prediction error, we keep both rules and replace their scores with measured finite-horizon returns. Simulating both modes from cloned decision states with their map inputs is computationally expensive, so we use 10 episodes per scene. T2-up remains 0/10 and T5 1/10.

  • Relative-return comparison. In T5, some withdrawals occur even after the current visual return exceeds the blind return. Earlier low visual scores remain in the averaging window, because the rule compares a smoothed visual score with the current blind score. Those different time references can cause withdrawal on a correct map.
  • Blind-return decline. This rule requests blind control when the current blind score is at or below its recent mean minus a margin. In T2-up, the decline rule still requests blind control when the visual score exceeds the current blind score both before and after smoothing. A falling blind score on difficult terrain can indicate a greater need for vision, not a map error.

The recorded decisions explain why better prediction alone need not resolve these failures. SURE's consistency criterion withdraws the visual residual when the map contradicts physical evidence, not simply when locomotion becomes harder.

AuthorsOpenReview ↗

Response to Reviewer r2Mf (4/4)

Q2(b). Miki-style belief adaptation

Answer to Q2(b): With the controller fixed, we compared adapting visual features with directly withdrawing an inconsistent visual correction. Belief adaptation achieves 99.98% on clean terrain and performs well on G1/G4, but achieves 0% on G2 false support versus 100% for SURE. Its recurrent encoder combines corrupted-map features with proprioceptive history and supplies adapted features to B+R. The encoder learns from clean and fault-matched rollouts, using clean-map actions and an auxiliary reconstruction objective.

A filled depression looks flat in the map. Both methods use proprioception, but belief adaptation uses it to adjust visual features, while SURE estimates physical support independently of the map. If adapted features retain false support, the fixed controller can continue to follow it. SURE checks the map against the support estimate and withdraws the inconsistent residual without first reconstructing visual features.

Q2(c). Matched-DR Visual

Answer to Q2(c): Both methods retain near-perfect clean success, but SURE's G1/G3 advantage persists after matching fault categories.

Matched-DR Visual trains for 30,000 PPO iterations. Each episode samples nominal maps (40%), flattening to omit protrusions or fill depressions (30%), map delay (20%), or downward shifts (10%). Delays range from 0.1 to 1.0 s in 0.1 s increments; shifts are uniform from 0.10 to 0.30 m. All episodes use uniform height noise in [−0.03, 0.03] m.

On clean terrain, SURE reduces velocity-tracking error, time-normalized first-order action-difference RMS, and torso angular velocity CVaR95 by 7.5%, 25.7%, and 36.7% relative to Matched-DR Visual. Both achieve 100% on G2, but SURE reduces tracking error by 15.7% and torso angular velocity CVaR95 by 24.1%. At equal G2 success, SURE tracks more accurately and reduces upper-tail torso angular velocity.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.