Search — mujoco

MuJoCo's MjSpec.to_xml() can change joint/state-vector ordering after exporting and re-importing modular MJCF using asset/model and attach. This breaks assumptions about stable state ordering across a round-trip.

MuJoCo's USD decoder can change inferred mass/inertia even when bodies have explicit inertials and there are visual-only geoms. This breaks MJCF->USD->MuJoCo validation that compares compiled inertials and deterministic rollouts.

Newton nut_bolt_sdf example fails QA because nuts move jitterily instead of slow and controlled under both xpbd and mujoco solvers. No stderr or crash occurs, but expected physical behavior is not met.

A previously disabled friction ramp test (mujoco_warp_cuda_0) needs follow-up to be re-enabled. The issue provides no additional details beyond the re-enable request.

The request is to validate Isaac Sim’s force-based conveyor formulation (Warp kernels applying friction-limited tangential forces from contact impulses) using Newton solvers. Newton’s current conveyor example uses a rotating belt mesh and ordinary friction and lacks behavior verification beyond stability.

MuJoCo's C model-editing API adds items one at a time and recomputes signature each time, which is painfully slow for large worlds. The request is for batch adding bodies/geoms/etc. to avoid repeated expensive recomputation.

MuJoCo’s USD decoder misses physics-purpose material bindings on colliders when different materials are used for visual vs physics purposes. This breaks contact property preservation in MJCF↔USD roundtrip validation.

MuJoCo’s USD decoder does not preserve several semantics needed for roundtrip validation: sites, model names, unlimited joints, and disabled-collider metadata. This undermines compiled-model comparisons and deterministic rollouts after import.

MuJoCo’s sample/compile cannot load USD decoder plugins unless the plugin is manually registered/loaded. Users want the sample to auto-load plugins from standard build/install directories before parsing.

In the mujoco_xpbd_coupled_solver example, the chain falls through the cube without interacting, indicating broken collision/constraint coupling. The provided repro uses newton[examples]==1.4.0.rc1 and Warp 1.15.0 on Windows.

In the mujoco_mpm_coupled_solver example, some particles appear stuck in mid-air and do not move. The repro is with newton[examples]==1.4.0.rc1 and Warp 1.15.0 on Windows.

The kamino_mujoco_admm_solver example consistently warns about truncated contacts. The log indicates this happens in a typical setup (Warp 1.15.0 on A40), suggesting defaults or limits are easy to exceed.

MuJoCo’s native island discovery uses quadratic temporary structures (ntree x ntree adjacency and column-index arrays) even for sparse incidence. Proposal is to remove the quadratic scratch by constructing connected components directly from constraint/tree incidence.

Newton viewers currently apply viewport-style USD purpose defaults (show default/proxy; hide guide/render) with no user override. Request is to add configurable purpose visibility so users can reveal guide geometry (e.g., MuJoCo sites) or hide other purposes.

In mujoco_vbd_coupled_solver with `--solver vbd`, there is excessive floor penetration. Rigid bodies and particles should not significantly penetrate the ground in a baseline configuration.

In mujoco_franka_vbd_cable_admm_solver, the Franka robot penetrates the floor. The example likely needs a placement adjustment and corresponding keypoint updates.

kamino_mujoco_admm_solver logs warnings that Newton rigid_contact_max exceeds Kamino model_max_contacts_host, causing contact truncation. Defaults should be aligned to prevent unexpected truncation in the example.

A previously tracked flaky GPU test failed again with intermittent 'nefc overflow' warnings suggesting increasing njmax. The recurrence indicates instability or nondeterminism in hydroelastic MuJoCo Warp CUDA testing.

GPU test test_mujoco_hydroelastic_penetration_depth_cuda_0 is flaky, failing with a ratio too low in one merge-queue run while later passing. The instability needs investigation to make outcomes reliable.

parse_usd's custom-frequency traversal does not honor ignore_paths while other traversals do. This causes prims that should be ignored to be processed during the custom-frequency pass.

Request to warn when CollisionPipeline auto-sizes a very large rigid contact buffer allocation. The warning should include the resolved slot count/bytes, drivers (pair count, world count), and remediation (set rigid_contact_max explicitly).

test_reset_functionality_elliptic_cuda_0 is flaky, intermittently hitting the solver's 100-iteration cap after resetting state. Failures occur in merge-queue GPU runs and correlate with a MuJoCo Warp solver upgrade noted in the report.

A MuJoCo Warp CUDA friction ramp test failed on an unrelated change, indicating intermittent behavior. The issue asks to investigate the cause of the flakiness in this GPU test.

MuJoCo's `mj_geomDistance` returns exactly 0.0 for separated convex mesh pairs under nativeccd, and the libccd fallback also violates a separating-plane lower bound. The issue is reproduced across multiple MuJoCo versions and persists on current main per the report.

Newton's SolverVBD ignores mu_torsional and mu_rolling material properties even though they exist in ShapeConfig and are used in other solvers like XPBD and MuJoCo. This creates visible behavioral differences for finite-radius objects like spheres/capsules in example scenarios such as a baggage conveyor.

Reinforcement learning (RL) for quadruped locomotion commonly depends on fixed, hand-crafted, and Markovian reward functions that limit both interpretability of learned policies and lack explicit control over gait behaviors. We introduce a framework where distinct gaits are specified using parameterized constraints expressed in Signal Temporal Logic (STL). These include safety bounds, gait synchronization constraints, command tracking, and actuation bounds. From these specifications, we develop a reward shaping mechanism that provides learning agents a dense, continuous reward landscape that encodes desired behavior. We define parametric STL templates for three speed regimes (walking-trot, trot, bound), calibrate their parameters from reference rollouts, and compute rewards from using smooth approximations of STL robustness over the rollouts. The generated rewards can be used to provide shaped gradients compatible with Proximal Policy Optimization (PPO). We instantiate the approach on Google's Barkour quadruped robot in MuJoCo XLA (MJX). We use parallelization within the simulator to improve training speeds and use domain randomization to robustify learned policies. We show that compared to a baseline of hand-crafted rewards, the STL-shaped rewards yield tighter velocity tracking and more stable training. Videos can be found on our project website: https://stl-locomotion.github.io/.

While robotic manipulation capabilities have advanced rapidly, physical safety remains a major barrier to deploying household robots: task success is insufficient if the robot damages itself or its surroundings. Simulation offers a harm-free alternative to costly and dangerous real-world training and evaluation, yet existing simulators lack general mechanisms to detect, quantify, and represent damage. To address this gap, we introduce OOPSIEVERSE, a unified simulation framework and benchmark for damage-aware household manipulation. OOPSIEVERSE provides damage as an explicit, physically-grounded, and taskagnostic signal by converting sources such as contact forces, temperature changes, and liquid interactions into corresponding mechanical, thermal or fluid damage. OOPSIEVERSE comprises two core elements: (1) DAMAGESIM, a simulator-agnostic framework for detecting and quantifying damage during navigation and manipulation, and (2) a suite of household tasks designed to evaluate common damage modes and distinguish between task completion and safe execution. We demonstrate the generality of our framework by instantiating DAMAGESIM in two simulators with different physics backends, OmniGibson (Nvidia Omniverse) and RoboCasa (MuJoCo). We further showcase the utility of OOPSIEVERSE across multiple use cases, including (1) guiding safer demonstration collection via real-time damage feedback, (2) learning safer manipulation policies through damage-conditioned imitation learning and reinforcement learning, (3) benchmarking the safety of state-of-the-art Vision Language Action policies, and (4) improving real-world safety of sim-to-real transferred policies. Together, our results highlight the potential of OOPSIEVERSE as an open-source foundation for systematic, scalable research on safe robot manipulation. For code and more information, please refer to https://robin-lab.cs.utexas.edu/oopsieverse/

Scalable reinforcement learning has popularized high-throughput sampling architectures, which significantly compresses the training time for off-policy methods in robotic locomotion. However, the rapid increase of data volume and update frequency undermines the stability of value-based methods and diminishes the plasticity of policy networks. To address these challenges, this work presents FastDSAC, a fast and high-performance variant of the Distributional Actor-Critic algorithm designed for parallel sampling scenarios. Specifically, we introduce a truncated Gaussian distribution to approximate the learned policy, which effectively excludes out-of-distribution actions that strain target value estimation while keeping necessary stochasticity for exploration. The proposed action constraint functions as an implicit regularization, which counteracts the plasticity loss typically caused by aggressive gradient updates. This preservation of network adaptability enhances sample efficiency, particularly in scenarios with a high update-to-data ratio, and accelerates the early training process. In contrast to prior fast reinforcement learning approaches that rely on discrete value distributions, our method utilizes a continuous Gaussian representation equipped with adaptive variance regulation, which improves value estimation accuracy by sampling confident and informative transitions. Extensive experiments on MuJoCo Playground and HumanoidBench demonstrate that FastDSAC not only stabilizes the overall training process but also achieves superior asymptotic performance and faster convergence compared to state-of-the-art baselines.