Robots trained entirely in simulation usually break down at the moment of contact. The simulator's model of pushing, sliding, and jamming is slightly wrong, so a policy that learned to press firmly in simulation presses far too hard, or at the wrong angle, on real hardware. Engineers currently respond on two fronts, both in software. They write reward terms that punish large forces and object interpenetration, and they calibrate or randomize contact parameters with great care. This proposal asks whether a cheap mechanical part can take over much of both jobs. The part is a 3D-printed wrist that stays stiff under normal loads and buckles into a soft state when contact force crosses a threshold. The policy being protected is a diffusion policy, a generator that produces a short chunk of future robot motions by gradually denoising random noise.
The core idea fits in two sentences. Diffusion policies commit to several steps of stiff position commands before they look at the world again, and this is exactly why they spike contact forces. A wrist that buckles at a set force caps that spike physically and instantly. If the wrist is modeled crudely in simulation and its parameters are randomized, then a policy trained with only a task-success reward and coarse contact randomization should transfer zero-shot to contact-rich tasks. In the baseline setup, those same tasks otherwise demand force-penalty rewards, penetration-aware training, or force-sensor feedback at high rates.
Core Claim
We claim that passive bimodal compliance changes what must be engineered in software. The change shows up in three places. The reward can drop its force, torque, and penetration penalty terms. Domain randomization, meaning the practice of training over a spread of simulated physical parameters so reality looks like one more sample, can be coarse over contact friction and stiffness, provided it spans the wrist's own uncertain parameters. And a position-chunked diffusion policy running at about 10 Hz can remain safe between replanning steps, with no force sensor in the loop.
The mechanism fits in one inequality. Let $F_b$ be the wrist's buckling force, $k_{low}$ its stiffness after buckling, $v$ the commanded end-effector speed, and $\Delta t_e$ the time a chunk executes before the next observation. The peak contact force during an unexpected collision is then roughly bounded by
$$F_{peak} \lesssim F_b + k_{low}\, v\, \Delta t_e .$$
A rigid wrist has no such cap, because its effective stiffness is huge. The bound follows from treating the buckled wrist as a soft spring loaded in series with the arm for at most one chunk's worth of overshoot. It is what makes open-loop chunks tolerable.
The multimodality matters for a specific reason. Compliance widens the set of approach strategies that succeed: a peg can be tilted in from several sides, and a box can be pivoted with the wrist loading against the wall in different ways. Demonstration data generated in simulation therefore becomes more multimodal, not less. A single Gaussian policy that averages those strategies would press into the middle, which is the one approach no expert used.
Prior Work and Where This Sits
Two papers motivate the proposal by showing the two halves of the problem in isolation. Adaptive Compliance Policy extends diffusion policies to output a reference pose, a compliance direction, and a stiffness. With this, it reaches 96% success on item flipping, against 14% for a stiff position policy and 23% for fixed compliance. It learns from real force-sensed human demonstrations and runs active admittance control. We keep the diffusion policy but move compliance into hardware and move data collection into simulation.
BiFlex shows the buckling wrist tolerating 50 mm of vertical error while keeping wiping forces under 15 N; a rigid wrist exceeded 15 N at 14 mm. Its control is hand-programmed, and it has no simulation model. We supply the learned policy and the simulator model it lacks. BiFlex also reports that TPU-95A relaxes under sustained load and oscillates. We treat that as uncertainty to randomize, not as a nuisance to ignore.
Three sim-to-real assembly systems constrain the design. They also set the bar, because they show how much software engineering rigid contact currently requires. Factory provides fast GPU contact simulation with signed distance fields (a representation of distance to an object's surface), and we reuse it directly. IndustReal adds penetration-aware policy updates, volumetric distance rewards, a curriculum, and an action integrator on top, reaching 83–99% zero-shot real success over 600 trials. Its remaining failures are wedging and slipping inside a rigid gripper, which is precisely what compliance targets. Zhang et al. tune admittance gains online with a force sensor and reach 90–100% at 0.05 mm clearance. None of the three uses a passive compliant element or a multimodal policy.
Schaff et al. offer the most direct evidence that compliant bodies forgive crude simulation: their linear-elastic soft crawler matched real displacement within 1 cm. Their control, however, is open-loop and has no observations. DrEureka automates both reward writing and randomization ranges for rigid robots, and its rewards lean on torque and acceleration penalties. We test whether such penalties become unnecessary with a buckling wrist, or even harmful because they suppress the deflections the wrist relies on.
DADP supplies an optional speed tool: biasing the diffusion prior with an inferred domain embedding allowed one-step sampling. We note that its assumption of static domain parameters may be violated by wrist stress relaxation. Diffusion Policy supplies the base architecture and its receding-horizon chunking, which runs at up to 10 Hz.
Research Program
The first step builds a cheap simulator model of the wrist. Inside Factory's Franka environments, the BiFlex-style wrist becomes a 6-DoF spring-damper between flange and gripper. The spring is piecewise-linear, stiff below the buckling force $F_b$ and soft above it, with hysteresis between loading and unloading. This model is fitted to bench measurements of the printed wrist. Deliberately, we do not use finite element simulation. The randomization distribution covers $F_b$, both stiffnesses, damping, and a slow relaxation term. We fit the ranges from repeated bench tests across several printed copies, so that print-to-print variation and fatigue are inside the distribution.
The second step generates multimodal demonstrations in simulation with a minimal reward. For each task we train several reinforcement-learning experts with PPO and only a sparse success term plus a distance-to-goal term. The experts are seeded with different initial strategies, such as tilt direction or pivot side. We also run a scripted planner. Their rollouts, collected under randomization, form the demonstration set. Contact friction and part stiffness are randomized over deliberately wide, uncalibrated ranges.
The third step distills these rollouts into a single diffusion policy that sees images and joint positions. It predicts 16-step position chunks and executes 8 before replanning. No force-torque input is used. If latency matters, we compare standard multi-step sampling against DADP-style prior biasing.
The fourth step analyzes where the gains come from, so the result is a measurement and not just a system. Sim-to-sim tests train under one contact model and test under shifted friction and stiffness, which traces a "tolerance envelope" for each hardware and reward configuration. Real tests then check whether that envelope predicts real transfer.
Baselines
The naive reference point is the same diffusion pipeline with a rigid wrist, trained on the same task-only reward and coarse randomization, then deployed zero-shot. We expect it to jam or trip the robot's force limits. It establishes how much of the gap the wrist closes.
The strongest existing method for assembly is IndustReal. We run its released pipeline on the same IndustRealKit pegs and gears, the same Franka, and the same success criterion. It gets its full reward and training machinery, and the same number of simulation steps as our experts. Its reported bar is 83–99% zero-shot real success across 600 trials (IndustReal). Beating it means one of two outcomes. Either we match its success within the confidence interval of about 100 trials per task while using only the sparse reward and none of its penetration-aware updates, or we exceed it on off-nominal part poses, where rigid wedging dominates its failures.
For force-sensitive manipulation, the strongest competitor is ACP, whose reported figure is 96% success on item flipping (Adaptive Compliance Policy). ACP learns from real demonstrations, which is an advantage we do not have. We therefore report its number as context and also reproduce a sim-trained version: diffusion over pose, compliance direction, and stiffness, with a simulated force sensor, on a rigid wrist under active admittance control. That version gets the same demonstrations and tuning budget as ours. Matching the sim-trained ACP without a force sensor would count as a win.
For tight insertion, the dynamic gain-tuning method of Zhang et al. reports 90–100% success at 0.05 mm clearance. We run it at our clearances as the active-compliance reference.
Three ablations isolate the key ingredient. The first mounts the real wrist but omits it from simulation. The second models the wrist but does not randomize its parameters. The third replaces the bimodal spring with a single linear spring matched in average stiffness. The last one separates the effect of buckling from the effect of simply being soft.
Why not just…?
Why not just run a rigid arm with software impedance control? This is a serious alternative. ACP's fixed-compliance baseline reached only 23% on item flipping, but that was trained from human demonstrations, so it does not settle our setting. We run fixed and learned impedance on the rigid arm at the same 10 Hz policy rate with the same simulated data. The bound above predicts that software compliance loses on force spikes during chunks, because its correction is limited by sensing and control delay. A buckling wrist responds mechanically. If impedance control at the robot's native low-level rate matches our peak forces and success, the wrist's contribution shrinks to "no force sensor needed," and we will report it that way.
Why not just keep a Gaussian PPO policy and add the wrist? This is the biggest threat to the diffusion half of the claim. Diffusion Policy's 95% versus 20% on real Push-T shows averaging can be devastating, but only when demonstrations are genuinely multimodal. We deploy the PPO experts directly with the wrist and measure strategy diversity in the demonstration set. If single-expert PPO with the wrist matches the diffusion policy, we conclude that the wrist, not the policy class, carries the result.
Why not just bolt the wrist onto a policy trained with a rigid wrist? This is our first ablation. Our guess is that it helps against collisions but hurts precision, because the gripper's deflection under load is invisible to the policy. If it matches the full method, then simulating the body is unnecessary, and the proposal's recommendation becomes simpler.
Why not just engineer the software pipeline more carefully, as IndustReal does? IndustReal already works. Our claim is not higher peak success on nominal insertion but equal success with fewer reward terms and coarser calibration. We quantify that by counting reward terms and sweeping reward weights, and we say plainly that a loss to IndustReal on nominal insertion is a possible outcome. The wrist's sub-1-cm deflection under stiff loads may be too loose for sub-millimeter clearances. If that happens, we report it as the regime where morphology does not substitute for software.
Decisive Experiments
We use three real tasks on a Franka Panda. The first is peg and gear insertion from the IndustRealKit, at the kit's clearances and with deliberately perturbed part poses. The second is ACP-style item flipping against a wall. The third is wiping over BiFlex-style uneven surfaces with 0 to 50 mm of height variation. We report success rate, peak and 95th-percentile contact force from an external force plate used only for evaluation, and jamming rate, with at least 50 trials per condition per task.
The central factorial crosses rigid versus modeled buckling wrist with task-only reward versus a force-penalized reward in DrEureka style, and runs both at coarse and at calibrated contact randomization. The proposal succeeds if the buckling wrist with task-only reward and coarse randomization is statistically indistinguishable from the calibrated, penalized configuration, while the rigid wrist loses substantially when calibration or penalties are removed.
A second check tests whether acceleration and torque penalties reduce success with the buckling wrist. It does this by measuring how often the policy drives the wrist past $F_b$ on tasks where buckling is useful. If they do reduce success, automated reward writers need to know the hardware.
The third check tests the force bound directly. We vary chunk execution length $\Delta t_e$ and approach speed $v$ and verify that measured peak force scales with their product on the buckling wrist but not on the rigid one.
The final check compares the sim-to-sim tolerance envelope for each configuration against its real success rate. This tests whether the cheap simulated measurement predicts transfer, which would let future hardware choices be screened before anything is printed.
References
- Adaptive Compliance Policy: Learning Approximate Compliance for Diffusion Guided Control
- BiFlex: A Passive Bimodal Stiffness Flexible Wrist for Manipulation in Unstructured Environments
- Factory: Fast Contact for Robotic Assembly
- IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality
- Bridging the Sim-to-Real Gap with Dynamic Compliance Tuning for Industrial Insertion
- Soft Robots Learn to Crawl: Jointly Optimizing Design and Control with Sim-to-Real Transfer
- DrEureka: Language Model Guided Sim-To-Real Transfer
- DADP: Domain Adaptive Diffusion Policy
- Diffusion Policy: Visuomotor Policy Learning via Action Diffusion