KinDERBench: Robotics Benchmark Suite

• KinDERBench is the evaluation component of the KinDER framework that systematically benchmarks physical reasoning in robotic learning and planning.
• It integrates 13 reference baselines spanning classical planning, RL, imitation, diffusion policies, and foundation-model approaches to address kinematic and dynamic challenges.
• The suite employs rigorous protocols with metrics like success rate, cumulative reward, and inference time, and validates performance via sim-to-real transfer experiments.

KinDERBench is the evaluation and benchmarking suite within the KinDER (Kinematic and Dynamic Embodied Reasoning) framework, developed to enable systematic comparison of physical reasoning performance in robot learning and planning. KinDERBench consists of an integrated set of protocols, metrics, and 13 reference baselines, spanning classical task-and-motion planning, reinforcement learning (RL), imitation learning with diffusion policies, generative planners, and foundation-model-based approaches. Its design supports unified and rigorous assessment of robotic agents on procedurally generated environments targeting core kinematic and dynamic reasoning challenges, decoupled from perception and language understanding. The KinDER framework and all associated pipelines, including KinDERBench, are fully open-sourced (Huang et al., 28 Apr 2026).

1. Evaluation Protocols and Methodology

KinDERBench employs standardized protocols to ensure reproducibility and comparability across algorithms and environments. For every environment–baseline pair, evaluations are performed using 5 random seeds, 50 episodes per seed, and a maximum episode length $T_{\max}$ (environment-specific). Planning baselines (e.g., Bilevel Planning, LLMPlan) utilize skill and concept predicates exposed in the KinDER API, whereas learning-based baselines can access either object-centric states or RGB+state observations.

Key parameters include:

• Randomized initialization: Each run samples task instances, object counts, and poses via procedural generation.
• Action and state tracking: Object-centric state dictionaries and vectorized actions, conforming to the KinDERGym API.
• Sparse reward structure: $r_t = -1$ at all timesteps except at first goal satisfaction, where $r_{T^*} = 0$. Protocols strictly separate training and test environments for methods requiring offline training.

2. Metrics

KinDERBench supports consistent multi-metric evaluation, implemented as follows:

1. Success Rate (SR):

$S = \frac{1}{N_{\rm ep}}\sum_{e=1}^{N_{\rm ep}} \mathbf{1}[\text{episode $e$ succeeded}]$

2. Cumulative Reward (Rwd):

$$R = \frac{1}{N_{\rm suc}} \sum_{e:\text{suc}} \sum_{t=0}^{T_e} r_t$$

where $r_t = -1$ for $t < T_e$, $r_{T_e} = 0$.

3. Inference Time (Inf-Time):

$$T_\text{inf} = \frac{1}{N_\text{ep}}\sum_{e=1}^{N_\text{ep}}\bigl(\frac{\text{wallclock}_{e}}{T_e}\bigr)$$

measured as seconds per step or per episode.

4. (Optional) Collision Frequency: Measured as total collisions divided by total timesteps.

These metrics emphasize generalization, efficiency, and robustness in physically grounded reasoning tasks (Huang et al., 28 Apr 2026).

3. Implemented Baselines

KinDERBench includes 13 baselines, representing major algorithmic paradigms in robot learning and planning:

Group	Baseline	Approach (brief)
Planning	BP	Bilevel Planning: symbolic+geometric planning, hand-coded skills, h_{FF}-guided search, high engineering cost
Foundation	LLMPlan, VLMPlan	Zero-shot (LLMs/VLMs), skill-sequence extraction, open-loop rollout
Foundation	LLMCon, VLMCon	In-context LLM/VLM with sample tasks, interpretable plan extraction
MPC	MPC	Model-Predictive Control using ground-truth physics
MBRL	MBRL	Learned dynamics model (two-layer MLP) + MPC
Diffusion	GSC	Generative Skill Chaining using diffusion over skill-latents
RL	PPO, SAC	Model-free RL (actor–critic MLPs, dense/sparse rewards)
Imitation	DP, DPES	Diffusion Policy (UNet backbone, DDIM, horizon-limited), state-augmented variant
Vision-Action	VLA	Pretrained VLA, fine-tuned on 100 KinDER demonstrations

Every baseline is implemented with precisely reported hyperparameters, architectures, and objective functions, including the PPO clip objective, SAC soft actor, and diffusion denoising loss. Skills and concept predicates are utilized through the KinDERGym Python library for planning algorithms (Huang et al., 28 Apr 2026).

4. Empirical Performance and Findings

Performance results are quantified as mean SR, Rwd, and Inf-Time across environments. Key findings reveal that:

• Bilevel Planning (BP) outperforms all other methods in success rate (SR $\sim$0.57 average), achieving near-perfect performance in basic kinematic scenarios with moderate environment complexity (e.g., SR=1.0 on Motion2D, Table3D, Shelf3D), but suffering steep computational costs on combinatorial, long-horizon, or dynamic environments (e.g., SR drops to 0.01 on DynPushPullHook2D with Inf-Time $r_t = -1$070 s).
• Foundation-model planners (LLMCon/VLMCon) demonstrate nontrivial reasoning (SR $r_t = -1$10.43), especially when provided context, outperforming zero-shot variants.
• Model-predictive control (MPC) with ground-truth physics surpasses model-based RL (MBRL), confirming difficulties with learned model accuracy (SR MPC=0.32, MBRL=0.08).
• Imitation and diffusion-policy baselines achieve meaningful performance on certain long-horizon tasks but lack general robustness (DP SR=0.24, DPES SR=0.25).
• Model-free RL approaches (PPO, SAC) underperform on most environments, unable to address sparse-reward, long-horizon requirements (PPO SR=0.13, SAC SR=0.02).
• Pretrained VLA achieves notable skill on tool-use environments, outperforming DP/DPES on tasks such as DynPushPullHook2D.

These results indicate that physical reasoning—particularly long-horizon manipulation under kinematic and dynamic constraints—remains an open challenge, with large gaps between classical planning and current learning-based methods (Huang et al., 28 Apr 2026).

5. Generalization and Analysis

KinDERBench includes systematic generalization tests. For example, out-of-distribution evaluation on DynObstruction2D—training on one obstacle and testing on up to three—shows stepwise decline in SR (e.g., DP: 0.35 → 0.26; VLA: 0.52 → 0.40) as task complexity increases.

Bilevel planning's scalability is assessed in StickButton2D, with SR dropping from 0.99 to 0.0 and planning time escalating from 0.85 s to 38.4 s as the number of buttons grows from 1 to 10. This highlights the computational bottlenecks in discrete-geometric search under combinatorial constraints.

A plausible implication is that current symbolic and learning paradigms require architectural or algorithmic advances to effectively scale to complex, dynamic, multi-object reasoning situations (Huang et al., 28 Apr 2026).

6. Real-to-Sim-to-Real Validation

KinDERBench is integrated with a real-to-sim-to-real pipeline for assessing sim–real transfer. In the Shelf3D task, the evaluation protocol involves:

1. Overhead vision-based detection for centroid and scale estimation,
2. Proprioceptive measurement of robot base and joint states,
3. MuJoCo twin initialization,
4. Motion plan generation (bilevel planner),
5. Real-robot waypoint execution using inverse kinematics and base control.

Domain randomization via Gaussian noise wrappers is used to test robustness. Quantitative results indicate close correspondence between simulation and deployed real-robot performance—e.g., Sim SR=0.90, Real SR=0.85, Position RMSE=0.02 m for Shelf3D—demonstrating relatively high sim-to-real fidelity for moderate-complexity object packing tasks (Huang et al., 28 Apr 2026).

7. Significance and Use Cases

KinDERBench, as the evaluation arm of the KinDER suite, establishes a unified, extensible, and highly controlled testbed for advancing research on physical reasoning in robotics. By encompassing a cross-section of algorithmic paradigms, standardized metrics, and real-world validation, it enables:

• Rigorous comparison across classical planning, RL, imitation, generative, and foundation-model methods,
• Isolation of specific kinematic and dynamic reasoning challenges, disentangled from perception and language,
• Rapid benchmarking for new methods in embodied physical reasoning,
• Assessment of generalization, sim–real transfer, and robustness under environmental variability.

KinDERBench is intended as a research accelerant, providing clear terrain on which to “reason about what to reason about” in embodied robotic intelligence (Huang et al., 28 Apr 2026).