Robot Tactile Olympiad (RoTO) Benchmark

• RoTO is a standardized benchmark suite that evaluates reinforcement and self-supervised learning for blind, tactile-based dexterous manipulation.
• It focuses on proprioceptive and binary contact inputs, intentionally excluding vision to challenge contact-rich, sensor-driven control tasks.
• The open-source platform supports large-scale experiments with GPU-parallelized simulations and standardized metrics for tasks like finding, bouncing, and Baoding.

The Robot Tactile Olympiad (RoTO) is a standardized benchmark suite for evaluating reinforcement learning (RL) and self-supervised learning (SSL) methods in dexterous robotic control using tactile sensing and proprioception. RoTO systematically excludes vision and privileged object state, focusing on "blind" manipulation and challenging contact-rich tasks that require skill beyond what is achievable by low-dimensional or visual-only feedback. RoTO and its successor, roto 2.0, serve as open, reproducible platforms supporting large-scale experimentation, multi-robot generalization, and robust baseline comparison in tactile-based RL research (Miller et al., 24 Oct 2025, Miller et al., 20 May 2026).

1. Motivation and Benchmark Design

RoTO addresses the long-standing fragmentation in tactile RL research, which was previously hampered by bespoke hardware, proprietary sensors, and limited, over-saturated orientation tasks (Miller et al., 20 May 2026). The benchmark’s core goal is to standardize the evaluation of RL and SSL methods for dexterous manipulation using only proprioceptive feedback and sparse binary contact data, thus prioritizing algorithmic advances over environment-specific engineering.

Key design features include:

• Evaluation exclusively through proprioception and per-link binary contacts—omitting privileged information such as vision or ground-truth pose.
• Open-source environments and tuned baselines to encourage reproducibility and lower the barrier to entry.
• Multiple robot morphologies for cross-platform generality (v2.0: Shadow Dexterous, Shadow Lite, Allegro, and ORCA hands; 16–24 DOFs).
• GPU-parallelized Isaac Lab environments enabling high-throughput training on single-node hardware (up to 8,092 instances) (Miller et al., 20 May 2026).

2. Task Suite and Physical Implementation

RoTO’s environment suite encompasses contact-intensive tasks of increasing complexity, instantiated in Isaac Lab with parameterized robots and strict observation constraints (Miller et al., 24 Oct 2025):

Task	Robot	DOF/Contacts	Object(s)	Core Objective	Episode Horizon (steps/s)
Find	Franka Panda	9	1 × 70 mm sphere	Locate sphere in plane	300 (~5 s)
Bounce	Shadow/S.Lite/Allegro/ORCA	16–24/14–20	1 × 70 mm stress ball	Bounce ball in midair	600 (~10 s)
Baoding	Shadow/S.Lite/Allegro/ORCA	16–24/14–20	2 × 38 mm spheres	Rotate balls in palm	600 (~10 s)

Each robot employs per-link binary contact sensors, joint position/velocity feedback, and last-commanded actions. Observations are temporally stacked, typically k=4 steps, with input spaces ranging from 100–228 dimensions depending on robot and task (Miller et al., 24 Oct 2025, Miller et al., 20 May 2026).

3. Observation, Action, and Reward Formalisms

RoTO imposes a strictly "blind" information structure:

Observations

• Proprioceptive: Joint angles, velocities, error (desired–actual), last actions.
• Tactile: Binary contact status per link.
• Input vector: $$\mathbf{o}_t = [c_{t-k+1}, p_{t-k+1}, \ldots, c_t, p_t]$$
• History: $k$ stacked timesteps (typically $k=4$)

Example for Shadow Hand: $m=17$ contacts, $d=20$ proprioceptive, $N=4 \times (17+20)=148$ (Miller et al., 20 May 2026).

Actions

• Continuous joint increments: $\mathbf{a}_t \in \mathbb{R}^{A}$, $A$ = number of controllable joints.

Reward Functions

• Find: $r^{Find}_t = -\|x_{EE}(s_t) - x_{sphere}\|_2$
• Bounce: $r^{Bounce}_t = 10\,\mathbb{1}\{$contact after ≥5 no-contacts$k$0
• Baoding:

$k$1

Optional torque penalties and episode truncation on object loss are standard. These dense- or event-based rewards are tailored to provoke dexterous, contact-driven strategies rather than simple motion planning (Miller et al., 24 Oct 2025, Miller et al., 20 May 2026).

4. Benchmark Protocol, Evaluation, and Baselines

Training protocol standardizes on millions of simulated steps with robust random seed sweeps (≥5 per setting), and wall-time–equivalent training on GPU clusters (Miller et al., 24 Oct 2025, Miller et al., 20 May 2026). Core evaluation metrics:

• Find: Time-to-locate (distance <2 cm), success rate.
• Bounce: Mean bounces per episode (maximum = 100 in 10 s), bounce peak precision.
• Baoding: Mean number of 360° rotations per episode (human record ≈13 in 10 s).

Reported results include mean ± 1 std over seeds, direct translation of episode return to interpretable physical metrics, and informal significance assessment via non-overlapping confidence intervals (Miller et al., 24 Oct 2025, Miller et al., 20 May 2026).

Baselines include:

• RL-Only: proprioceptive (+ optional tactile) MLP agent via PPO.
• RL+SSL: hybrid architectures with auxiliary SSL objectives—Tactile Reconstruction (TR), Full Reconstruction (FR), Forward Dynamics (FD), and Tactile Forward Dynamics (TFD). Notably, FD and TR auxiliaries yield gains across all tasks (Miller et al., 24 Oct 2025).
• RL+SSL+Separated Memory: detaching SSL replay buffer from on-policy memory boosts sample efficiency, particularly in Baoding rotations.

Empirical results: RL+SSL agents with FD objectives achieve, for example (Baoding), 25 rotations per 10 s (vs. human ≈13; RL-only ≈5). Blind agents in roto 2.0 attain up to 13 rotations, a notable order-of-magnitude faster than prior tactile RL SOTA (Miller et al., 24 Oct 2025, Miller et al., 20 May 2026).

5. Large-Scale Implementation and Reproducibility

RoTO and its v2.0 release are tightly integrated with Isaac Lab for high-throughput, GPU-resident simulation (Miller et al., 20 May 2026). Key features:

• Batch training: 4,096–8,092 parallelized environments for training, 100 for evaluation.
• Single-GPU sufficiency: Full-scale experiments run on high-end NVIDIA hardware (Titan V, V100).
• Policy and simulation loop entirely on GPU, eliminating CPU-GPU transfer bottlenecks.
• Fully open-sourced codebase, configuration, Docker compatibility, and hyperparameter sweep machinery; installation includes cloning, workspace setup, and YAML/gRPC-based task launching.

These frame RoTO as a leading testbed for tactile-RL reproducibility and cross-lab benchmarking at scale.

6. Connection to Related Benchmarks and Extensions

RoTO bridges the gap between vision-dominated RL and high-fidelity contact-based reasoning, supporting compositional evaluation across morphologies and tasks. TACTFUL (Kamtikar et al., 23 Jun 2026) demonstrates that dense taxel-based hands and transformer policies enable robust real-world tactile exploration and identification, with 77.8% success and ≈1.5 cm Chamfer-L2 reconstruction error, validating tactile-driven approaches in object localization and manipulation.

Extensions discussed include:

• Integration of richer tactile sensor modalities (resistive arrays, force/torque, etc.).
• Additional tasks: grasping, object identification, texture and compliance probing, dynamic and multi-object scenarios.
• Multi-modal fusion (e.g., sparse vision, acoustic, proprioceptive cues).
• Light-weight sim-to-real transfer, curriculum learning, and hybrid observation models.

A plausible implication is that future RoTO challenges will incorporate more varied sensor types, real-world hardware loops, and complex object interaction regimes, building upon both simulation and experimental work presented in TACTFUL (Kamtikar et al., 23 Jun 2026).

7. Recommendations and Outlook

For optimal benchmarking and research on RoTO:

• Prefer sparse binary contacts and proprioception; privilege-free inputs yield more generalizable, realistic results.
• Joint RL+SSL training with forward-dynamics or tactile-reconstruction auxiliaries is recommended, using separated replay buffers.
• Hyperparameter sweep and rigorous reporting (≥5 seeds, mean±std, physical metrics) are essential.
• Source code, environment configs, and training scripts should be open for reproducibility.
• Initial focus on simple tactile formats is advocated; richer modalities can be phased in as needed.

RoTO and its extensions are positioned to unlock progress on core challenges in tactile-based RL, offering a scalable, objective foundation for comparative evaluations and future innovations in robotic manipulation (Miller et al., 24 Oct 2025, Miller et al., 20 May 2026, Kamtikar et al., 23 Jun 2026).