A robot hand that rotates an object with one carefully optimized finger pattern fails completely when one finger stalls, a fingertip pad wears smooth, or the object turns out heavier than expected. Modern contact-rich controllers recover gracefully from small errors. They cannot recover from structural ones, because they only ever hold one strategy. We propose to precompute, offline, an archive of many good manipulation strategies. Each strategy is indexed not by how the motion looks but by which physical contacts it depends on: which fingertips, palm regions, or environment surfaces carry load, and whether those contacts stick or slide. When a fault occurs, the robot switches to a stored strategy whose dependence footprint avoids the failed resource, instead of re-solving from a configuration where its local optimizer is already stuck.
The method is MAP-Elites, a quality-diversity algorithm that keeps the best solution found in each cell of a grid over behavior descriptors. The new part is the descriptor. Existing quality-diversity robotics organizes diversity around kinematics: gait contact ratios, velocities, joint-angle embeddings, trajectory shapes. Strategies that look different can still lean on the same finger and therefore fail together. Our claim is that redundancy in manipulation should be measured as failure independence, and that a descriptor built from the contact forces a trajectory optimizer already computes gives this independence nearly for free.
Core Claim and Positioning
Three groups of prior work motivate the proposal. The first shows that repertoires buy robustness. Scaling MAP-Elites to Deep Neuroevolution evolves archives of neural gaits and recovers from leg damage by picking a different elite. Quality-Diversity Actor-Critic learns a single skill-conditioned network with successor-feature diversity critics on Brax locomotion. From Tabula Rasa to Emergent Abilities (URSA) discovers quadruped skills on hardware and selects among them after damage with Iterative Trial and Error. One Solution is Not All You Need (SMERL) and Discovering Diverse Nearly Optimal Policies with Successor Features keep several near-optimal locomotion policies under a return constraint. All of this work is locomotion or navigation, and all of it uses kinematic descriptors.
URSA's authors also note a problem that constrains our design. Because their skills share one network, damage to one actuator can degrade many skills at once. Behavioral redundancy is supposed to prevent exactly this. We therefore store elites as separate trajectory-plus-feedback plans rather than as codes into one shared network, and we test that choice directly.
The second group supplies the machinery. Dexterous Contact-Rich Manipulation via the Contact Trust Region (R-CTR) is a model-predictive controller (one that repeatedly re-optimizes a short plan from the current state) that forbids unphysical pulling forces at contacts. It reaches 2.0 mm translation error against 8.6 mm for an ellipsoidal trust region, and it optimizes Allegro Hand policies in under ten minutes on a laptop CPU. Contact-Implicit Model Predictive Control for Dexterous In-hand Manipulation pairs a 20 Hz contact-implicit planner with a 100 Hz tactile compliance layer. Quasistatic contact-rich manipulation via LCQP solves contact as a complementarity program, meaning a contact either has zero gap or zero force, never both. It does this in 7 ms per solve.
These solvers output contact forces and stick/slide states as a by-product. That is exactly the information our descriptor needs, and it makes them sample-efficient variation operators in place of black-box mutation.
The third group shows why these solvers are not enough alone. Katayama et al.'s controller falls from 100% to 48–60% success under noise and model error. Contact Mode Guided Motion Planning argues that local controllers get trapped in contact equilibria and need a higher-level choice of contact mode. Each of these systems commits to one locally optimal path.
The closest neighbors are the generative multimodal policies. Learning Multimodal Behaviors from Scratch with Diffusion Policy Gradient clusters trajectories by their shape. Behavioral Mode Discovery keeps up to ten discrete modes during diffusion fine-tuning on ManiSkill. Both measure diversity in trajectory space, not in contact dependence, and neither is evaluated on actuator or contact faults.
The Descriptor
For a candidate strategy $\pi$ and a set of contact resources $r = 1,\dots,R$ (the four fingertips, the palm, and any environment surface), let $\Lambda_r(\pi)$ be the total normal impulse resource $r$ transmits over the episode. The descriptor is
$$ b_r(\pi) = \mathbb{1}\left[\frac{\Lambda_r(\pi)}{\sum_{r'} \Lambda_{r'}(\pi)} > \tau\right], \qquad b_{R+1}(\pi) = \text{fraction of loaded contact time spent sliding}, $$
where $\tau$ is a load-share threshold (default 0.1). The binary part gives at most $2^R$ cells, 64 for six resources, so the archive stays small without hand-picked continuous features. The sliding fraction is binned coarsely into three bins. The intended use is simple: a strategy whose footprint has $b_r = 0$ should be largely unaffected by a fault on resource $r$.
Whether this holds is an empirical question and the central thing we test. For example, a finger with zero load share may still need to move out of the way, so a locked joint could still block it.
Research Program
Archive construction. We build the archive in simulation using the quasidynamic contact model that R-CTR uses, on two task families. The first is in-hand cube reorientation on the Allegro Hand, between the 1,000 start/goal configuration pairs used by Suh et al. The second is non-grasp pivoting and cylinder manipulation in Katayama et al.'s PyBullet setting.
Each MAP-Elites iteration selects an elite and applies a footprint mutation, which forbids or newly permits one resource by adding a zero-normal-force constraint on it. It then refines the result with R-CTR trajectory optimization, reads the descriptor off the solver's own contact forces, and inserts the result if it beats the cell's incumbent on task cost. Fitness is goal error plus effort. We keep an elite only if it is within a fixed tolerance of the best cost for its start/goal pair, mirroring the near-optimality constraint of SMERL and DSP.
Deployment. At run time the nominal elite executes under closed-loop R-CTR tracking with a compliance layer. A stall detector triggers a switch when progress toward the goal stops or tracking error persists. Selection then works in one of two ways. With a fault guess from joint-current and tactile residuals, we pick the best elite whose footprint excludes the suspected resource. Without a guess, we run Iterative Trial and Error over the archive as URSA does, using the footprint to set the prior correlations between cells.
Fault suite. We perturb the system in four ways: a disabled finger (zero torque), a locked joint, fingertip friction reduced by half (a worn pad), and object mass doubled. We also include Katayama et al.'s sensor noise and model error. The fault suite stays out of archive construction, so faults are unseen at deployment.
Baselines
The naive reference is the strongest single-strategy controller with no archive. For reorientation this is R-CTR closed-loop MPC from Suh et al., run with their published settings. For pivoting it is the relaxed LCQP controller from Katayama et al., whose reported success under noise and model error is 48–60%. We expect both to hold up under noise but to fail under structural faults. On structural faults such as a disabled finger, no reported result found on Lacuna for R-CTR or LCQP, so we measure it ourselves.
The strongest competitor, and the real threat, is fault-informed online replanning. When a stall is detected, R-CTR is restarted from the current state with K random initializations, and it is also given the same zero-force constraint on the suspected resource that our selector uses. This gives it the diagnostic information our method gets. It is the strongest competitor because it differs from our method in only one respect: whether distinct strategies are found offline from good starting states or online from the stuck one. We run it at matched wall-clock latency and at ten times our latency, on the same simulator, fault suite, and stall detector.
We beat it only if success after the fault is higher at matched latency, specifically by at least 15 absolute points on disabled-finger and locked-joint faults. If it matches us only once given ten times our latency, the archive's value is speed rather than reachability, and we will report it that way.
The first ablation keeps the identical MAP-Elites and R-CTR pipeline but swaps in kinematic descriptors: object-pose trajectory features plus fingertip path clusters under Dynamic Time Warping, in the style of DDiffPG. This isolates the descriptor. Both archives get equal evaluation counts and are pruned to equal size.
The second ablation stores the footprint archive as one skill-conditioned network distilled from all elites, in the style of QDAC and URSA. This isolates modular storage against limb dependency.
Finally we train a learned single policy with heavy domain randomization over friction, mass, and actuator strength. We give it a simulator-sample budget equal to the total budget of archive construction. It is tuned on the randomization ranges but never on the held-out faults.
Why not just…?
Why not just replan online with a contact-implicit MPC? This is the most serious alternative, and the restart baseline above is built to test it. The evidence against it is indirect. Cheng et al. report that local planners get trapped in contact equilibria. Katayama et al.'s success drops to 48–60% under model error even without structural faults. A replanner that starts from a jammed configuration may only find solutions near that jam, while archive elites start from configurations where alternative footprints were feasible.
We also run a third variant that warm-starts R-CTR from the archive elite. If that hybrid beats pure lookup, we will report the contribution as "archive as warm-start" rather than as a replacement for replanning.
Why not just train one robust policy with domain randomization? Robust Contact-rich Manipulation through Implicit Motor Adaptation reports that domain randomization is too conservative and is beaten in final state error on Hit, Push, and Reorientation tasks. SMERL shows that a single mode cannot cover perturbations that demand mutually exclusive solutions, and a disabled finger is such a case. Our matched-budget domain-randomization baseline tests this directly on our fault suite.
Why not just estimate the fault and adapt one policy online? The same implicit-adaptation paper finds that explicit point estimation of physical parameters is unreliable under intermittent, noisy contact and loses to probabilistic adaptation. Our method does not need an accurate estimate: Iterative Trial and Error over footprint cells works with no diagnosis at all. We report results with diagnosis and without it.
Why not just use an existing kinematic quality-diversity method? The kinematic-descriptor ablation answers this. If it matches our fault coverage, the contact-footprint idea is unnecessary and the paper's claim fails.
Proposed Experiments and Decisive Checks
All metrics are averaged over the held-out start/goal pairs and five archive seeds. The primary metric is success after the fault. Secondary metrics are time to recovery and the number of trials Iterative Trial and Error needs.
A third metric is archive fault coverage: the fraction of (fault, task) pairs for which at least one elite succeeds. We measure it by exhaustively testing every elite under every fault offline. This separates "the archive contains a solution" from "the selector found it."
The decisive check on the core claim is a failure-correlation analysis. For each pair of elites we compute the correlation of their success indicators across faults, and we compare the footprint archive with the kinematic archive at equal size. The proposal stands only if the footprint archive has clearly lower average pairwise failure correlation and higher fault coverage. We also check whether $b_r = 0$ predicts survival of a fault on resource $r$, which tests the descriptor's intended use. If it does not, for example because locked fingers block motion even without carrying load, we extend the footprint to include joint-space workspace occupancy and test again.
A hardware stage follows only if the simulation checks pass. It transfers the reorientation archive to an Allegro Hand and repeats the disabled-finger fault, using Suh et al.'s 150 consecutive transitions as the nominal reference.
References
- Scaling MAP-Elites to deep neuroevolution
- Quality-Diversity Actor-Critic: Learning High-Performing and Diverse Behaviors via Value and Successor Features Critics
- From Tabula Rasa to Emergent Abilities: Discovering Robot Skills via Real-World Unsupervised Quality-Diversity
- One Solution is Not All You Need: Few-Shot Extrapolation via Structured MaxEnt RL
- Discovering Diverse Nearly Optimal Policies with Successor Features
- Dexterous Contact-Rich Manipulation via the Contact Trust Region
- Contact-Implicit Model Predictive Control for Dexterous In-hand Manipulation: A Long-Horizon and Robust Approach
- Quasistatic contact-rich manipulation via linear complementarity quadratic programming
- Contact Mode Guided Motion Planning for Quasidynamic Dexterous Manipulation in 3D
- Learning Multimodal Behaviors from Scratch with Diffusion Policy Gradient
- Behavioral Mode Discovery for Fine-tuning Multimodal Generative Policies
- Robust contact-rich manipulation through implicit motor adaptation